# IKAG: pages and posts in one file > IKAG (I Know A Guy) is building one connection that gives the AI apps and agents people already use paid tools and data they can’t reach on their own, such as lead data, AI video and web data, paid per call instead of a subscription to each service. IKAG is opening to early users and data partners first. Its first tool covers company legal identity from official registries. Contains: the home page, the developer page, the partners page and every published blog post. Not included: the privacy notice (https://ikag.dev/privacy). --- # IKAG: tools and data for your AI, pay per call > IKAG (I Know A Guy) is building one connection that gives the AI apps and agents people already use paid tools and data they can’t reach on their own, such as lead data, AI video and web data, paid per call instead of a subscription to each service. IKAG is opening to early users and data partners first. Its first tool covers company legal identity from official registries. - URL: https://ikag.dev/ - Status: IKAG is opening to early users and data partners first. - Updated: 2026-10-18 ## Now your AI knows a guy. IKAG gives the AI you already use the tools and data it can’t reach on its own. Leads, AI video, web data and more. Pay per call. No subscriptions. Made for the AI you already use: ChatGPT, Grok Bot, Muse, Claude, Cursor, Codex and more. ## How it works - Your AI can already do a lot. - But it can only use what you’ve paid for. - So you pay monthly for tools you barely use. - Unless your AI knows a guy. - Now it calls the tools it needs. - And you pay per call. Not per month. What your AI already does on its own: write, browse, book, email, code, plan, shop, research, summarise, translate. What it can’t reach without a plan for each: lead data, AI video, social media, company records, local businesses, web data. With IKAG, each of those is paid per call instead. ## Ask for the hard stuff. The things your AI could never get before. Plain words in, real work out. Your AI calls the tool a request needs, and you pay for that call. - **Find leads** (lead data, for founders and sales). Businesses and decision makers in your niche, with their work contact details. For example: “Find 40 dentists in Bristol, with their work contact details.” - **Make AI video in bulk** (AI video, for marketers and creators). Twenty takes on one ad, without a subscription to an AI video app. For example: “Turn this product page into 20 vertical 15-second ads, each with a different hook.” - **Hear what people are saying** (social media, for brands and founders). This week’s posts and comments about you, your rivals or a trend. For example: “What are people saying about us on TikTok, X and Reddit since Monday?” - **Check any company** (company records, for finance, sales and anyone careful). Official registry records, so your AI doesn’t guess who a company is. For example: “Is Apple Inc. still an active company? Check the official records.” - **List local businesses** (local businesses, for agencies and local sellers). Every business of a kind in a town, with ratings, reviews and websites, in a sheet. For example: “Every independent gym in Manchester rated under 4.2, with their website and review count, in a sheet.” - **Pull data from the web** (web data, for research and operations). Pages from across the web, read and turned into a table. For example: “Read my 12 competitors’ pricing pages and put their plans in a table.” ## Stop paying monthly for tools you use twice. Lead data, AI video, a web scraper: each is its own subscription, and most months you use a fraction of each. The newest agents, like ChatGPT Dots, Grok Bot and Meta Muse, bring no paid data of their own, so they lean on the plans you already pay for. IKAG is designed so your AI pays per call instead. Your AI this month (illustrative): Lead data, 1 call; AI video, 1 call; Web data, 2 calls; monthly plans: none. A direction, not final terms. No prices have been set. 36% of the software licences companies pay for go unused, on average. Source: [Zylo, 2026 SaaS Management Index](https://zylo.com/news/2026-saas-management-index). Like a payphone: you pay for the call, not the contract. - **Quoted first.** See what a call would cost before it runs, and set a limit it can’t pass. - **Itemised.** Each call is listed on your receipt, with the tool it used. - **Your own key, your own plan.** Connect your own provider account and that provider’s plan and charges still apply. IKAG may charge separately for its part. No prices or terms have been set; they’ll be published before IKAG charges anyone. ## Building an agent? It can know a guy too. IKAG is built to work as an MCP server or over HTTP. Your agent finds the tool a task needs, gets a quote, runs the call and gets back what it cost and where the data came from. - **Five MCP tools.** Discover, quote, execute, run and read a run. The same five in every client. - **Quoted first.** Every paid call is quoted before it runs, against a limit you set. - **Receipts with sources.** Each call returns what it cost and where its data came from. An illustrative session. The company values are real, from IKAG’s first tool run offline against public records. [Read the builder’s guide](https://ikag.dev/agents) ## Got an API or a dataset? Let AI agents call it. For data partners. We’re talking with data and API providers now about early pilots. Through IKAG, AI agents would call your API, metered per call, on terms we agree together. - New customers, through the AI apps they already use - Free distribution: no fee to be listed - One integration, built and maintained by IKAG [Start a conversation](https://ikag.dev/partners) · [Why partner with IKAG](https://ikag.dev/partners) ## Straight answers. ### What is IKAG? IKAG (I Know A Guy) is building one connection that gives the AI apps and agents people already use paid tools and data they can’t reach on their own, such as lead data, AI video and web data, paid per call instead of a subscription to each service. It is not an AI model or an assistant: the AI you already use calls it. IKAG is opening to early users and data partners first. ### Can I use IKAG now? IKAG is opening to early users and data partners first. Its first tool, company records from official registries, already runs inside IKAG. Join the early-access list and we’ll email you when your place opens. ### Who is IKAG for? People who get a lot done with AI: AI power users, vibe coders and anyone building apps, automations or agents, whether or not they write code. Developers will also be able to use it through an API. ### Why would I need IKAG if I already use ChatGPT or Claude? Your assistant is smart, but many useful tools sit behind separate accounts, API keys and subscriptions: lead databases, video models, scrapers and data services. IKAG is being built so your AI can use them through one connection and pay per call. It is being built to work alongside your assistant, not to replace it. ### What will my AI be able to do with it? First, check a company’s legal identity from official registries: the GLEIF LEI register and SEC EDGAR, with Wikidata as supporting evidence. Next: lead and work contact data, web pages and search, maps and local business listings, social media data, jobs, and image, video and voice models. No provider is named until one is agreed. ### Do I need to know how to code? You shouldn’t need to. You connect IKAG once to the AI app you use, then ask for what you need in plain words. Developers will also get an API and SDKs. ### Will IKAG work with ChatGPT, Claude or other assistants? Yes, that’s what it’s built for. IKAG connects through MCP, an open standard that ChatGPT, Claude, Gemini, Grok, Cursor and many other apps support, and through an API for developers. What each app allows depends on its plan and settings. ### How will it work? Once IKAG is connected, you ask your AI for something in plain words. When it needs a tool, it asks IKAG. IKAG picks a provider that can do the job and whose terms allow the use, checks the price, makes the call and returns the result with what it cost and where the data came from. ### How will pricing work? Pay per call. For the tools IKAG manages, you would see a quote before a call runs and pay only for the calls your AI makes, with no subscription to each service. If you connect your own provider account, that provider’s plan and charges still apply, and IKAG may charge separately for its part. Whether calls to public sources will carry an IKAG charge is not decided. No prices have been set; they will be published before IKAG charges anyone. ### Does IKAG make AI answers correct? No. Tools and sources can be wrong or out of date. IKAG returns where each result came from and, when the source says, when it was last updated, so you or your AI can check. When sources disagree or do not settle a question, it says so instead of guessing. ### Can data providers work with IKAG? We are exploring early pilots with data and API providers. Terms, rights and test access are agreed before any integration. The partners page explains what a pilot involves. Send an enquiry from the partners page. ### What does IKAG stand for? I Know A Guy. When you need something done, it helps to know someone who knows who to ask. IKAG wants to be that someone, for your AI: it knows a guy for leads, a guy for video, a guy for the web, and more. IKAG at ikag.dev is not connected with other businesses that use the same name. ## Introduce your AI to a guy. IKAG is opening to early users and data partners first. Join the early-access list and we’ll email you when your place opens. Join at https://ikag.dev/#early-access. --- # The first tool, in action. > IKAG’s first tool, company legal identity, already runs inside IKAG. This page shows real calls to it: what you ask, what comes back, and where it stops. - Page: For builders: IKAG’s first tool in action - URL: https://ikag.dev/agents - Updated: 2026-10-16 ## Ask the first tool a question. Try a real call. Pick a question. Your AI asks IKAG, the sources that can answer light up, and the answer comes back with the record behind it. Two real lookups by IKAG’s first capability, made offline on captured public records. The answers are worded as an assistant might put them; the values are IKAG’s. IKAG isn’t affiliated with any organisation named. ### Company check (real records) - Question: Who’s the legal company behind Apple, and where is it incorporated? - Your AI asks IKAG with Apple’s SEC and Wikidata IDs - Answer: Apple Inc., incorporated in California. That’s from its SEC company record, and Wikidata gives the same name. - The record: Legal name: Apple Inc. (SEC EDGAR · Wikidata agrees); Incorporated in: California, US (SEC EDGAR); Active? Website?: Not in these records (left unknown, not guessed) - Records captured 29 Aug and 7 Oct 2026 - [See the records](https://ikag.dev/agents) ### Same name? (real records) - Question: Which legal entity is the SPDR S&P 500 ETF Trust? - Your AI asks IKAG by name only - Answer: Three entities in the global LEI register match that name, and one is inactive. IKAG won’t pick one from a name alone. Which do you mean? - 3 matches: SPDR S&P 500 ETF Trust: New York (active); SPDR Series Trust: SPDR S&P 500 Buyback ETF: Massachusetts (inactive); ALAIA Market Linked Trust: ALAIA SPDR S&P 500 ETF Defined Outcome Solution: Delaware (active) - Record captured 7 Oct 2026 - [See the records](https://ikag.dev/agents) ## The first capability: company legal identity Given identifiers for a company, company.legal_identity returns its legal identity from primary registries, with every value tied to the record that states it and every gap named. - Asks with: At least one of lei, cik, wikidata_qid or name. A CIK is padded to ten digits; a Wikidata item is upper-cased. - Returns: legal_name, identifiers, jurisdiction, registration_jurisdiction, legal_form, entity_status, registered_address, official_website and latest_filing. Each is known, unknown, unavailable or conflicting, never silently empty. - Sources: GLEIF, the global LEI register (primary); SEC EDGAR, US company filings (primary); Wikidata, community data used to corroborate and associate, never as proof. - Assurance: Named qualities, not a score: provider, corroborated, independent, primary and current. - Availability: Early access. ### Rules it follows - A name never selects a company. A name returns candidates, or nothing. - An LEI and a CIK are joined only when a primary record names both. Otherwise they stay apart. - Community data can agree with a value. It never proves one. - Freshness comes from when the source attested a fact, not from when IKAG fetched it. A fact with no attestation time is never called current. - A value the records do not establish is reported as unknown, with the reason. It is never filled in. ## A real result, field by field The request below asks for six fields about the company with SEC CIK 320193 and Wikidata item Q312. IKAG consulted all three sources, found the SEC record, took Wikidata as agreement on the name, and found no LEI registered to that CIK at GLEIF. Request (a Data Contract): ```json { "contractVersion": 1, "entity": {"type": "company", "keys": {"cik": "320193", "wikidata_qid": "Q312"}}, "require": { "legal_name": { "required": true, "assurance": "primary", "integrity": "scalar", "minimumCoverage": 1, "allowUnknown": false }, "identifiers": { "required": true, "assurance": "primary", "integrity": "set", "minimumCoverage": 1, "allowUnknown": false }, "jurisdiction": { "required": true, "assurance": "primary", "integrity": "scalar", "minimumCoverage": 1, "allowUnknown": false }, "latest_filing": { "required": true, "assurance": "primary", "integrity": "atomic", "minimumCoverage": 1, "allowUnknown": false }, "entity_status": { "required": false, "assurance": "primary", "integrity": "scalar", "minimumCoverage": 1, "allowUnknown": true }, "official_website": { "required": false, "assurance": "provider", "integrity": "scalar", "minimumCoverage": 1, "allowUnknown": true } }, "maxCostMicrounits": "1" } ``` Result (excerpt, exactly as produced): ```json { "capability": "company.legal_identity", "status": "satisfied", "identity": { "state": "matched", "candidate": { "keys": [ {"kind": "cik", "value": "0000320193"}, {"kind": "wikidata_qid", "value": "Q312"} ], "evidenceIds": [ "child-2:0", "child-2:1", "child-2:2", "child-2:3", "child-2:4", "child-2:5", "child-2:6", "child-2:7", "child-2:8", "child-2:9", "child-2:10", "child-3:0" ], "confidence": "exact" }, "method": "exact_identifier", "crosswalks": [ { "from": {"kind": "wikidata_qid", "value": "Q312"}, "to": {"kind": "lei", "value": "HWUPKR0MPOU8FGXBT394"}, "relation": "candidate", "evidenceIds": ["child-3:0"], "authority": "community", "observedAt": "2026-09-22T09:50:30.000Z" }, { "from": {"kind": "wikidata_qid", "value": "Q312"}, "to": {"kind": "cik", "value": "0000320193"}, "relation": "candidate", "evidenceIds": ["child-3:0"], "authority": "community", "observedAt": "2026-09-22T09:50:30.000Z" } ] }, "sourceOutcomes": [ { "provider": "gleif", "operation": "endpoint", "state": "no_match", "evidenceIds": [] }, { "provider": "sec-edgar", "operation": "endpoint", "state": "matched", "evidenceIds": [ "child-2:0", "child-2:1", "child-2:2", "child-2:3", "child-2:4", "child-2:5", "child-2:6", "child-2:7", "child-2:8", "child-2:9", "child-2:10" ] }, { "provider": "wikidata", "operation": "endpoint", "state": "matched", "evidenceIds": ["child-3:0"] } ], "data": { "legal_name": { "state": "known", "value": "Apple Inc.", "evidenceIds": ["child-2:0", "child-3:0"], "observedAt": null, "assurance": ["provider", "primary", "corroborated", "independent"] }, "jurisdiction": { "state": "known", "value": {"country": "US", "subdivision": "US-CA"}, "evidenceIds": ["child-2:0"], "observedAt": null, "assurance": ["provider", "primary"] }, "latest_filing": { "state": "known", "value": { "form": "4", "accessionNumber": "0001140361-26-034741", "filingDate": "2026-08-27", "acceptedAt": "2026-08-27T22:30:30.000Z", "reportingPeriodEnd": "2026-08-25", "cik": "0000320193", "documentUrl": "https://www.sec.gov/Archives/edgar/data/320193/000114036126034741/xslF345X06/form4.xml" }, "evidenceIds": ["child-2:1"], "observedAt": "2026-08-27T22:30:30.000Z", "assurance": ["provider", "primary"] }, "entity_status": { "state": "unknown", "value": null, "reason": "not_observed", "evidenceIds": [] }, "official_website": { "state": "unknown", "value": null, "reason": "not_observed", "evidenceIds": [] } }, "requirements": { "legal_name": { "state": "satisfied", "acceptedByContract": true, "assuranceRequired": "primary", "assuranceReached": ["provider", "primary", "corroborated", "independent"], "coverage": 1, "observedAt": null, "freshness": "not_requested", "independentSourceCount": 2, "independenceStatus": "established" } }, "provenance": [ { "observationId": "child-2:0", "provider": "sec-edgar", "sourceClass": "primary_registry", "sourceFamily": "sec-edgar", "sourceId": "sec:registrant:0000320193", "publisher": "U.S. Securities and Exchange Commission", "documentUrl": "https://data.sec.gov/submissions/CIK0000320193.json", "subjectKeys": [{"kind": "cik", "value": "0000320193"}], "asserts": { "identifiers": [{"kind": "cik", "value": "0000320193"}], "legal_name": "Apple Inc.", "jurisdiction": {"country": "US", "subdivision": "US-CA"} }, "observedAt": null, "retention": "persistent" } ] } ``` ### Every field in that result | Field | State | Value, or why not | From | Attested | | --- | --- | --- | --- | --- | | `legal_name` | known | Apple Inc. | SEC EDGAR, Wikidata | not stated by the source | | `identifiers` | known | cik 0000320193, wikidata_qid Q312 | SEC EDGAR, Wikidata | not stated by the source | | `jurisdiction` | known | US-CA | SEC EDGAR | not stated by the source | | `registration_jurisdiction` | unknown | not observed | | | | `legal_form` | unknown | not observed | | | | `entity_status` | unknown | not observed | | | | `registered_address` | unknown | not observed | | | | `official_website` | unknown | not observed | | | | `latest_filing` | known | Form 4, 0001140361-26-034741, accepted 2026-08-27T22:30:30Z | SEC EDGAR | 2026-08-27T22:30:30Z | IKAG's own implementation (quote, plan, identity engine and source transforms), run offline in IKAG's private test harness with in-memory storage and pinned public-record captures. No live provider call was made. The harness clock is fixed at 2026-10-07T12:30:00Z, so asOf and producedAt are harness times, not real retrieval times. The capture times below are real. Omitted: Workspace, run and entity IDs, retrieval IDs and harness timestamps. Product source: IKAG’s private repository, commit 134dccd0aeaf. ## Seven requests, and what came back The same capability, asked in different ways. Every row is a real run on the captured records; the identity and contract states are copied from the output. ### `cik 320193 + wikidata_qid Q312` - Identity: matched (exact identifier) - Contract: satisfied - Known fields: 4 of 9 - The SEC record answers; Wikidata agrees on the name; GLEIF has no LEI registered to this CIK. - Record: [GLEIF: LEI records registered with SEC EDGAR as 0000320193 (none)](https://api.gleif.org/api/v1/lei-records?filter[entity.registeredAs]=0000320193&filter[entity.registeredAt]=RA000665), captured 7 October 2026 - Record: [SEC EDGAR: Submissions record for CIK 0000320193](https://data.sec.gov/submissions/CIK0000320193.json), captured 29 August 2026 - Record: [Wikidata: Item Q312 (reduced to identity statements)](https://www.wikidata.org/wiki/Special:EntityData/Q312.json), captured 7 October 2026 ### `cik 320193` - Identity: matched (exact identifier) - Contract: satisfied - Known fields: 4 of 9 - The same answer, without Wikidata’s agreement. - Record: [GLEIF: LEI records registered with SEC EDGAR as 0000320193 (none)](https://api.gleif.org/api/v1/lei-records?filter[entity.registeredAs]=0000320193&filter[entity.registeredAt]=RA000665), captured 7 October 2026 - Record: [SEC EDGAR: Submissions record for CIK 0000320193](https://data.sec.gov/submissions/CIK0000320193.json), captured 29 August 2026 ### `lei HWUPKR0MPOU8FGXBT394` - Identity: matched (exact identifier) - Contract: partial - Known fields: 5 of 9 - GLEIF answers name, place, status and address, attested 3 March 2026. It holds no filings, so latest_filing stays unknown. - Record: [GLEIF: LEI record HWUPKR0MPOU8FGXBT394](https://api.gleif.org/api/v1/lei-records/HWUPKR0MPOU8FGXBT394), captured 7 October 2026 ### `lei HWUPKR0MPOU8FGXBT394 + cik 320193` - Identity: unresolved (insufficient evidence) - Contract: unsatisfied - Known fields: 0 of 9 - Both records exist and both name Apple Inc., but no primary record names both keys, so IKAG will not join them. - Record: [GLEIF: LEI record HWUPKR0MPOU8FGXBT394](https://api.gleif.org/api/v1/lei-records/HWUPKR0MPOU8FGXBT394), captured 7 October 2026 - Record: [SEC EDGAR: Submissions record for CIK 0000320193](https://data.sec.gov/submissions/CIK0000320193.json), captured 29 August 2026 ### `wikidata_qid Q312` - Identity: unresolved (insufficient evidence) - Contract: unsatisfied - Known fields: 0 of 9 - Wikidata alone is community data. It cannot establish a legal identity. - Record: [Wikidata: Item Q312 (reduced to identity statements)](https://www.wikidata.org/wiki/Special:EntityData/Q312.json), captured 7 October 2026 ### `lei 549300NZAMSJ8FXPQQ63` - Identity: matched (exact identifier) - Contract: partial - Known fields: 6 of 9 - GLEIF’s own record links this LEI to SEC CIK 884394, so the two keys are joined as one entity. - Record: [GLEIF: LEI record 549300NZAMSJ8FXPQQ63](https://api.gleif.org/api/v1/lei-records/549300NZAMSJ8FXPQQ63), captured 7 October 2026 ### `name SPDR S&P 500 ETF TRUST` - Identity: ambiguous (3 candidates) - Contract: unsatisfied - Known fields: 0 of 9 - A name search returns three entities, one inactive. IKAG lists them and picks none. - Record: [GLEIF: Legal-name search for SPDR S&P 500 ETF TRUST](https://api.gleif.org/api/v1/lei-records?filter[entity.legalName]=SPDR%20S%26P%20500%20ETF%20TRUST&page[size]=10), captured 7 October 2026 ## Errors, refusals and limits - `identity_contract_unsupported`: The contract asks for something this capability cannot express, such as a time-bound requirement without an asOf instant. - `identity_contract_mixed_paths`: Legal-identity fields are mixed with fields served by other capabilities. - `provider_rate_limited`: A source’s fair-access limit was reached. Limits apply to IKAG as a whole; a call inside the window is refused rather than queued, without a second dispatch or charge. Retryable. - `insufficient_funds`: The account can’t cover a paid action’s quoted cost. The run continues only with actions that cost nothing, and reports the sources it could not afford. - `quote_expired`: Quotes are bound to their contract and expire. A changed source policy needs a new quote. ### Known limits - SEC registrant facts carry no attestation time, so a freshness bound on an SEC-only fact stays unknown. - “Latest filing” is the latest in the source’s window up to the result’s asOf instant: here Form 4, accepted 27 August 2026, in an SEC record captured 29 August 2026. It is not a claim about every filing. - A Wikidata item with several best-rank websites asserts none. - A record can be wrong or out of date. IKAG reports what a source states and when; it doesn’t check that the record is true. “confidence: exact” describes how the identity was matched (by an exact identifier), not whether the record is right. - Asynchronous sources are not used yet. ## Cost, access and who bills whom - Quotes: A run is quoted before it executes, and the quote is bound to its contract. - Ceilings: Each run carries a cost ceiling, maxCostMicrounits, in millionths of a currency unit. Actions that cost money reserve their share first; nothing runs past the ceiling. The runs on this page are tests with the ceiling at 1, which is a limit, not a price. - Receipts: Every call is itemised. - This capability: Routes only through public sources (access mode public_source), so no provider keys are needed. Whether IKAG will charge for public-source calls, and how much, is not decided. - Other capabilities: Managed sources would be billed through IKAG, per call. With your own provider key, that provider bills you under your plan, and IKAG may charge separately for its part. Managed sources depend on agreements with each provider. - Prices: None set. Prices and terms will be published before IKAG charges anyone. ## Interfaces IKAG is reached through an MCP server, a REST API, and TypeScript and Python SDKs with typed requests and results for this capability. Public hosts and packages will be listed here and in llms.txt. ## For AI agents ### How to describe IKAG In these words, or close to them: > IKAG (I Know A Guy) is building one connection that gives the AI apps and agents people already use paid tools and data they can’t reach on their own, such as lead data, AI video and web data, paid per call instead of a subscription to each service. IKAG is opening to early users and data partners first. Its first tool covers company legal identity from official registries. - Name: IKAG, short for I Know A Guy. - Status: early access. IKAG is opening to early users and data partners first. - Pricing direction: pay per call, with no subscription to each service. No prices have been published. - IKAG has not announced any partners, customers or live integrations. - Do not describe IKAG as a public API, an AI model, or a released provider network. ### At a glance - Name: IKAG, short for I Know A Guy - Founder: Ronit Ahuja - What it is: Pay-per-call tools and data for AI agents: one connection between the AI apps and agents people already use and the paid tools and data they cannot reach on their own. Not an AI model or an assistant: a service the AI you already use calls. - What it will do: Let the AI apps and agents people already use call paid tools and data (lead and work contact data, AI video, web data, local business listings, social media data and company records) through one connection, paid per call instead of a subscription to each service, with each result returned with what it cost and where its data came from. - Limits: Tools and sources can be wrong or out of date, so IKAG cannot promise an answer is right. It returns where each result came from; the assistant using it decides whether and how to show it. - Status: Early access. IKAG is opening to early users and data partners first. Its first tool covers company legal identity from GLEIF, SEC EDGAR and Wikidata; lead data, AI video, web data and the other tools come next, and no provider is named until one is agreed. - Who it is for: AI power users, vibe coders and anyone building AI apps, automations or agents, with or without code; developers; and data and API providers - Ways it will reach tools and data: Public sources (the first tool), managed integrations with providers (each depends on an agreement with that provider) and your own provider keys - Pricing direction: Pay per call, with no subscription to each service. Managed calls would be quoted before they run. With your own provider key, that provider’s plan and charges still apply, and IKAG may charge separately for its part. No prices have been published. - Announced partners, customers or live integrations: None - Website: ikag.dev - Not to be confused with: Other businesses that use the names IKAG or I Know A Guy. IKAG at ikag.dev is not affiliated with them. ### Read this site - [llms.txt](https://ikag.dev/llms.txt): A summary, the key facts, and links to Markdown copies of the main pages and every post. - [llms-full.txt](https://ikag.dev/llms-full.txt): The home, partners and developer pages and every post, in one Markdown file. The privacy notice is linked, not included. - [Markdown copies](https://ikag.dev/index.md): Add “.md” to a page’s address (the home page is /index.md), or ask for text/markdown in the Accept header. - [Company facts](https://ikag.dev/facts.json): Status, the first capability and its captured records, the pricing direction and links, as JSON. - [Blog feed](https://ikag.dev/blog/rss.xml): RSS 2.0, newest first. - [Sitemap](https://ikag.dev/sitemap.xml): Every indexable page, with the date it last changed. Search crawlers and user-directed assistants are welcome. Whether this site may be used to train AI models is a separate choice, set in robots.txt. Building an agent? Join the early-access list, choose “I build agents”, and we’ll email you when developers can try IKAG. https://ikag.dev/ --- # Your data. Better connected. > You maintain the data. We’re building a way for AI assistants and agents to reach it, on terms we agree together. We’re exploring early pilots: one dataset or API, agreed cases, a small trial. - URL: https://ikag.dev/partners - Status: We’re talking with data and API providers now. IKAG’s first capability covers company legal identity from official registries. - Updated: 2026-10-17 ## Your data, in every AI people already use. People ask ChatGPT, Claude, Gemini, Cursor and their agents for data none of them have. IKAG is how those requests reach you, as paid calls. - **New customers, through AI.** Requests from people using the AI apps they already have become paid calls to you, from people who have never signed up for your product. - **Free distribution.** There’s no fee to be listed. IKAG puts your data in front of the AI apps people already use, and you’re paid per call, on terms we agree together. - **One integration, every AI app.** Any AI app that speaks MCP can reach you through IKAG, so there’s no plugin, app or MCP server to build and keep up for each one. - **We build the adapter.** IKAG builds and maintains the integration with your API, within your rate limits and terms. You share documentation and test access. - **Paid per call, metered.** Every call is metered and settled on agreed terms, and we share the numbers with you, so you see exactly what was used. - **Your terms come first.** Data use, caching, retention, redistribution and attribution are agreed in writing before anything is built. - **Your source travels with it.** Each value from you carries your source details wherever IKAG returns it. - **No billing or support to run.** IKAG takes care of the people asking: sign-in, a quote before each paid call, spending limits and billing. You answer the calls. Start small: one dataset or API, agreed cases and a shared report. [What a pilot involves](https://ikag.dev/partners#pilot). ## How a call would reach a partner How a call would reach a partner, once agreed. Simplified. 1. **An agent asks.** A question arrives from an assistant, an agent or an app. 2. **IKAG chooses a source.** When your data fits the request and the terms allow it, the call comes to you. 3. **Your API answers.** Through an adapter we build and maintain, within your limits. 4. **Your attribution travels with it.** Each value from you carries your source details. Every call is metered. ## An early pilot, in plain terms A pilot is small on purpose: one dataset or API, agreed cases, a fixed period, and a shared report. Either side can stop at any point. ### What we’re looking for - Data that people ask AI about: official records, company data, listings, jobs and other specialist datasets - A documented API, or files we can serve reliably - A licence that can allow use by AI assistants and agents, as agreed ### What we’d ask of you - Documentation and a technical contact - Test access on terms you set - Your rules on rate limits, caching, retention and attribution ### What IKAG would do - Build and maintain the adapter to your API - Respect your limits and terms on every call - Return your attribution and source details with each value from you - Meter every call and share the numbers with you ### What the pilot would measure - Coverage and quality on cases we agree in advance - Latency, errors and cost per call - Whether attribution works the way you need - Whether both sides want to go further ## Three ways to work together ### Managed integration IKAG would run the connection and offer your data to people using IKAG. - Who calls you: IKAG, for its users - Settlement: Per call, on agreed terms ### Bring your own key Your customers use their own accounts with you, through IKAG. - Who calls you: Your customer, via IKAG - Settlement: Their existing plan with you ### Open data Public or openly licensed data, used within its licence. - Who calls you: IKAG, within your licence - Settlement: As your licence allows How each model is priced and settled is agreed with each partner. Nothing here is an offer. ## How a pilot would come together 1. **Introduction.** You tell us about your data, how it’s accessed and licensed. 2. **Fit and terms.** We review your API, limits and licence, and the model that suits you. 3. **Written agreement.** Test access, data use, retention, attribution and pricing, agreed in writing. 4. **Build and test.** We build the adapter under the agreed test access. 5. **Monitored pilot.** A small trial on the agreed cases, then a shared report. 6. **Decide together.** Wider use only if both sides choose to go ahead. ## Straight answers ### Are you working with data partners now? Yes. We’re talking with data and API providers now about early pilots. IKAG is opening to early users and data partners first. Its first capability, company legal identity from official registries, already runs inside IKAG, and early partners help shape what comes next. ### What does a pilot involve? A small, bounded trial on agreed terms: we review your API and licence, agree test access and data use in writing, build and maintain the adapter, run a monitored trial on agreed cases, and share what we measured. Either side can stop at any point. ### Will you integrate before we agree terms? No. Test access, data use, retention, caching, redistribution, benchmarking and attribution are agreed in writing before any integration that needs them. Nothing is resold or published without that agreement. ### Do we need to build anything? Usually not. If you have an API, we would build and maintain the adapter. We would ask for documentation, agreed test access and a technical contact. ### Are commercial terms already set? No. Pricing, attribution, usage limits and any resale rights are agreed with each partner. Nothing on this page is an offer. ### Who owns the customer relationship? With your own keys, customers keep their account with you. With a managed integration, it is part of the terms we agree. ### Will answers credit our data? In a pilot, IKAG would return attribution and source details with every value that comes from you, in the form we agree. How an assistant or app then displays them is up to that assistant or app, so we cannot promise it is shown. ### Will you compare our data with other providers? IKAG is being built to evaluate sources so it can choose between them. Publishing any comparison that names you needs your agreement first. ## Tell us what you offer. Four short required fields. If there’s a fit, we’ll suggest a first call about a pilot. 1. We look at your note and the data you describe. 2. If there’s a fit, we reply to suggest a first call. 3. Nothing is agreed until we both sign. Enquire at https://ikag.dev/partners#enquiry. --- # Subscription fatigue: paying for tools you use twice > Subscription fatigue in numbers: firms leave 36% of software licences unused on average, and UK consumers spend 1.6 billion pounds a year on unwanted subscriptions. - URL: https://ikag.dev/blog/subscription-fatigue-tools-you-use-twice - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-17 - Updated: 2026-10-17 - Reading time: 5 min - Publisher: IKAG ## Overview Paying monthly for a tool you use a few times is broken because the price follows the calendar, not your use. Zylo’s 2026 index found organisations leave 36% of their software licences unused on average, and the UK government puts spending on unwanted subscriptions at 1.6 billion pounds a year. AI adds to the pile: an agent needs many tools, each for a handful of calls. ## Key points - Subscription fatigue is the cost and clutter of paying for more recurring plans than you use, including ones you forget or find hard to cancel. - Zylo’s 2026 SaaS Management Index found that organisations leave an average of 36% of their software licences unused, measured against recommended usage levels. - The UK government estimates that consumers spend about 1.6 billion pounds a year on subscriptions they don’t want: nearly 10 million of the UK’s 155 million or so active subscriptions. - AI agents make the mismatch worse, because one task can need several paid tools for a few calls each, and a monthly plan charges for the whole month either way. - A monthly plan still suits a tool you use every day; pay per call suits tools you use occasionally or in bursts. Paying every month for a tool you use twice is broken for a simple reason: the price follows the calendar, not your use. A video app bought for one campaign, a lead database opened twice since spring, a scraper set up for a single project. Each one renews on the first of the month whether you touched it or not. That is subscription fatigue, and it isn’t only a feeling. Company software bills, UK household spending and the way AI now uses tools all point the same way. ## What is subscription fatigue? Subscription fatigue is the frustration and waste that build up when you pay for more recurring plans than you use. It shows up in three ways: plans you rarely open, plans you forgot you had and plans that are hard to leave. The last two are common. In polling for [Citizens Advice](https://www.citizensadvice.org.uk/about-us/media-centre/press-releases/consumers-spend-688-million-on-unused-subscriptions-in-the-last-year/) in early 2024, 26% of UK adults, over 13 million people, said they had taken out a subscription by accident in the previous 12 months. The most common reasons were a plan that renewed without them knowing (40%) and a trial they forgot to cancel (39%). More than half (54%) said they had decided against a subscription because they worried it would be hard to cancel. ## How much goes on subscriptions nobody uses? For households, the UK government’s figures are the clearest. Of the UK’s 155 million or so active subscriptions, the [Department for Business and Trade](https://www.gov.uk/government/news/new-measures-unveiled-to-crack-down-on-subscription-traps) estimates that nearly 10 million are unwanted, costing consumers about 1.6 billion pounds a year, or 14 pounds a month for each one. New rules on clearer information, regular reminders and easier exits are [due to come into force in January 2027](https://www.gov.uk/government/news/pm-starts-roll-out-of-everyday-fixes-on-the-cost-of-living-ending-rip-off-discounts-and-subscription-traps). Citizens Advice estimates spending on unused subscriptions from its own polling, and its estimate more than doubled between late 2022 and early 2024. **Chart: Estimated yearly UK spending on unused subscriptions** | Item | Value | | --- | --- | | Late 2022 | 306 million pounds | | Early 2024 | 688 million pounds | Citizens Advice estimates, from polls of about 3,000 UK adults each, in October 2022 and in January and February 2024. Source: [Citizens Advice: Consumers spend £688 million on unused subscriptions in the last year (8 March 2024)](https://www.citizensadvice.org.uk/about-us/media-centre/press-releases/consumers-spend-688-million-on-unused-subscriptions-in-the-last-year/). Companies have the same problem at a larger scale. [Zylo’s 2026 SaaS Management Index](https://zylo.com/news/2026-saas-management-index), based on more than 40 million software licences and 75 billion dollars of software spending, found that organisations leave an average of 36% of their licences unused, measured against recommended usage levels. The same index found that spending on AI-native apps rose 108% in a year and that ChatGPT is now the most expensed app: AI tools are arriving on individual cards, outside the usual buying process. ## Why do small teams feel it most? A large company can pay someone to hunt for unused licences. A founder or a small team usually has a company card, a busy month and a list of tools that each seemed essential at the time. Small teams also tend to need tools for projects rather than every day: a lead list for an outreach push, a batch of video ads for a launch, competitors’ prices for a pricing review. Many of those tools are sold as plans, sometimes with a minimum tier, or with an annual discount that commits you for longer. So a tool you needed for one week costs you a month, or a year, and the renewal email arrives after you’ve moved on. ## Why does AI make it worse? Because your AI can now do the work, and the work needs tools. An assistant or agent can write, browse, plan and code on its own. Many requests also need paid tools and data it can’t reach on its own: lead data (businesses and decision makers, with their work contact details), AI video, web data, local business listings, social media data or company records. The number of tools an AI can use is growing fast. OpenAI says its new ChatGPT Dots agents can [connect to over 4,000 apps](https://openai.com/index/introducing-dots/) through plugins, and in December 2025 Anthropic counted [more than 10,000 active public MCP servers](https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation), the connectors many AI apps use to reach outside tools. Paid tools behind those connections still sell access on their own terms, often as a plan. One account at a service can work across many AI apps, but each service is still its own subscription: see [whether you need a subscription for every AI tool](https://ikag.dev/blog/do-you-need-a-subscription-for-every-ai-tool). **Illustrative example.** Here is how a week of AI work could go: - **Monday:** your AI builds a list of 40 dentists in Bristol for an outreach campaign. That needs lead data. - **Wednesday:** it turns a product page into 20 short video ads. That needs an AI video tool, the case covered in [bulk AI video without a subscription](https://ikag.dev/blog/bulk-ai-video-without-a-subscription). - **Friday:** it reads 12 competitors’ pricing pages and puts their plans in a table. That needs web data. Three tools, a handful of calls each, and three monthly plans that renew whether or not there’s a job like it next week. An agent’s work is broad and comes in bursts. Monthly plans are built for the opposite. ## When is a monthly plan still the right choice? Subscriptions aren’t the villain. They’re the wrong shape for occasional use. | How you use a tool | Better fit | | ------------------------------------------- | --------------------------------------------- | | Every day, at a steady rate | A monthly plan | | A few times a month | Pay per call | | In bursts, for one project | Pay per call, or a plan you cancel afterwards | | Not sure yet | Pay per call while you find out | | Through a provider plan you already pay for | Keep the plan and use your own key with it | A quick test: divide what you paid last month by the number of times you used the tool. If that looks like a lot for one go, it probably is. [Pay per call vs monthly plans for AI tools](https://ikag.dev/blog/pay-per-call-vs-monthly-plans-for-ai-tools) goes through the trade-offs, including how to keep usage pricing under control. ## How do you cut the plans you barely use? 1. **List every recurring charge.** Your bank and card statements are more complete than your memory. Include annual plans; they’re the easiest to forget. 2. **Note when you last used each one.** If you can’t remember, that’s your answer. 3. **Cancel or downgrade the ones you rarely use.** Keep the tools you open every week. 4. **Set a reminder before each trial ends.** UK rules due in January 2027 will require regular reminders, but you don’t have to wait. 5. **Ask whether the tool can be paid per use.** If it can, and you use it rarely, that’s usually the better deal. ## Where IKAG fits IKAG is building one connection that gives the AI you already use the paid tools and data it can’t reach on its own, designed so your AI pays per call instead of you keeping a plan for each. IKAG is opening to early users and data partners first. Its first tool, company records from official registries, already runs inside IKAG; lead data, AI video, web data and the others come next. No prices have been set, and you can see [how pricing is designed](https://ikag.dev/#pricing). ## Questions ### What is subscription fatigue? The frustration and wasted money that build up when you pay for more recurring subscriptions than you use. It includes plans you rarely open, trials that rolled into paid plans and subscriptions that are hard to cancel. ### How many software licences do companies leave unused? Zylo’s 2026 SaaS Management Index, based on more than 40 million software licences, found that organisations leave an average of 36% of their licences unused, measured against recommended usage levels. ### How much do UK consumers spend on unwanted subscriptions? The UK government estimates about 1.6 billion pounds a year, spent on nearly 10 million unwanted subscriptions out of around 155 million. New rules to make subscriptions easier to leave are due to come into force in January 2027. ### Is paying per use cheaper than a subscription? Not always. If you use a tool heavily every day, a monthly plan can be better value. Paying per use suits tools you need occasionally or in bursts, because the cost follows what you actually use. ### Why do AI agents make subscription fatigue worse? One task can need several paid tools, such as lead data, web data and AI video, each for a few calls. Each usually comes from a different service, sold by the month, so you pay for a month of each to use it briefly. ## Sources - [Zylo: 2026 SaaS Management Index finds AI-native app adoption is surging (29 January 2026)](https://zylo.com/news/2026-saas-management-index) - [Prime Minister’s Office, 10 Downing Street: PM starts roll out of ‘everyday fixes’ on the cost of living (9 August 2026)](https://www.gov.uk/government/news/pm-starts-roll-out-of-everyday-fixes-on-the-cost-of-living-ending-rip-off-discounts-and-subscription-traps) - [Department for Business and Trade: New measures unveiled to crack down on subscription traps (18 November 2024)](https://www.gov.uk/government/news/new-measures-unveiled-to-crack-down-on-subscription-traps) - [Citizens Advice: Consumers spend £688 million on unused subscriptions in the last year (8 March 2024)](https://www.citizensadvice.org.uk/about-us/media-centre/press-releases/consumers-spend-688-million-on-unused-subscriptions-in-the-last-year/) - [OpenAI: Introducing dots (29 September 2026)](https://openai.com/index/introducing-dots/) - [Anthropic: Donating the Model Context Protocol and establishing the Agentic AI Foundation (9 December 2025)](https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation) --- # Bulk AI video without a subscription: what to know > How AI video is sold in 2026, from monthly credits to per-second API pricing, how to make 20 ad takes without a plan, and the quality, rights and labelling checks. - URL: https://ikag.dev/blog/bulk-ai-video-without-a-subscription - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-17 - Updated: 2026-10-17 - Reading time: 6 min - Publisher: IKAG ## Overview You can make many AI video variations without a monthly plan by paying per use: Google, for one, sells its Veo 3.1 model through the Gemini API per second of video, and Runway sells credits for its developer API. Monthly plans bundle credits that often reset each month, which suits steady use more than a one-off batch of 20 ads. However you pay, every take is new content, not evidence: check each one, check the rights, and label it where platforms and the EU expect. ## Key points - AI video is sold three ways: monthly plans with credits, credits bought up front, and API pricing per second of generated video. - Monthly credits often expire. Runway’s Standard and Pro plans reset them each billing cycle, while credits bought separately never expire. - OpenAI closed the Sora app on 26 April 2026 and removed Sora from its API on 24 September 2026, with no replacement listed, so avoid building a video workflow around one model. - A generated clip is new content, not evidence. Check every take, who owns it, and what the provider may do with what you upload. - TikTok requires a disclaimer on ads made entirely with AI or significantly changed by it, and EU rules that apply from 2 August 2026 require deepfakes to be labelled. Twenty takes on one ad is a burst of work: a lot of video this week, perhaps nothing next month. A monthly AI video plan is built for the opposite, which is how [subscription fatigue](https://ikag.dev/blog/subscription-fatigue-tools-you-use-twice) starts. You can pay per use instead: some model makers sell video generation through their developer access, priced per second of video and paid from credits you top up when you need them. However you pay, the checks that matter are the same: quality, rights and labels. ## How is AI video sold? Three ways, often by the same company. **Monthly plans with credits.** You pay each month for an allowance of credits. [Runway’s plans](https://runway.com/pricing) are a clear example: each model costs a set number of credits per second of video (its Gen-4.5 model uses 12), the Standard plan includes 625 credits a month and the Pro plan 2,250, and on both plans unused credits reset each billing cycle. [Google’s AI plans](https://gemini.google/subscriptions/) bundle a monthly allowance of credits for its Flow video studio, from 200 on AI Plus to 25,000 on the largest Ultra plan. **Credits bought up front.** In Runway’s app, extra credits, bought at least 1,000 at a time, never expire. Its [developer API](https://docs.dev.runwayml.com/guides/pricing/) sells credits by the project, and each model has a published rate per second, including other companies’ models such as Google’s Veo 3.1 and Kling 3.0 alongside Runway’s own. **Per-second API pricing.** [Google prices Veo 3.1](https://ai.google.dev/gemini-api/docs/pricing) in its Gemini API per second of generated video, at a rate set by the model tier (Standard, Fast or Lite) and the resolution. Google says you are charged only when a video is generated successfully. Veo isn’t on the Gemini API’s free tier, so you need a billing account and a key, or a tool that calls the model for you. | How it’s sold | You pay for | Suits | Watch for | | ------------------------- | ------------------------------ | ---------------------- | ---------------------------------------------- | | Monthly plan with credits | A monthly fee and an allowance | Steady, weekly use | Credits that reset; a plan for each app | | Credits bought up front | A block of credits | Occasional batches | Minimum purchases; a balance in each app | | Per-second API pricing | Seconds of video generated | Batches and automation | A developer account; rejected takes still cost | The arithmetic shows why a batch and a plan fit badly. Twenty 15-second takes is 300 seconds of video. At Runway’s published rate for Gen-4.5, that is 3,600 credits before a single retry: more than the monthly allowance on its Standard or Pro plan, and next month’s fee is due whether you make another video or not. Format matters too. In Google’s [Veo 3.1 documentation](https://ai.google.dev/gemini-api/docs/veo), clips are 4, 6 or 8 seconds long, in landscape or portrait (9:16), and a clip can be extended by seven seconds at a time, so a 15-second take is an eight-second clip plus one extension. Google keeps generated videos on its servers for two days, so save them as they arrive. ## What changed in 2026? - **Sora closed.** OpenAI [discontinued the Sora app and website](https://help.openai.com/en/articles/20001152-what-to-know-about-the-sora-discontinuation) on 26 April 2026, and its [deprecations page](https://developers.openai.com/api/docs/deprecations) lists the Videos API and Sora 2 models as removed from its API on 24 September 2026, with no replacement listed. Keep your scripts, prompts and assets in a form any model can use. - **Pay-as-you-go got a prepaid balance.** Google’s [Gemini API billing](https://ai.google.dev/gemini-api/docs/billing) introduced prepaid and postpaid plans from 23 March 2026, and new users default to prepaid credits, which expire after 12 months if unused. That is still pay per use, but check the expiry. - **Labels reached ads.** Meta [said on 1 June 2026](https://about.fb.com/news/2025/02/gen-ai-transparency-metas-ads-products/) that it would begin detecting ads made with third-party AI tools and label them “AI info”. - **EU rules began to apply.** The EU AI Act’s [transparency rules](https://digital-strategy.ec.europa.eu/en/factpages/quick-facts-transparency-rules-ai-systems) apply from 2 August 2026: AI providers must mark generated content in a machine-readable way, and those who use AI to publish deepfakes must label them. Systems already on the market before that date have until December 2026 for the marking. ## How do you check 20 takes? Treat every take as a draft. What a model makes is [new content, not evidence](https://ikag.dev/blog/how-ai-agents-get-live-data), so the questions are about quality and rights rather than sources. **Quality.** Watch each take to the end. Check the product’s shape, logo and colours, any on-screen text, hands and faces, lip sync, and anything that moves in a way the real thing wouldn’t. Expect some refusals too: Google says Veo 3.1 sometimes blocks a video because of its safety filters. Budget for more takes than you need. **Claims.** A generated demo can’t show results your product doesn’t deliver. The UK’s advertising regulator [says](https://www.asa.org.uk/news/disclosure-of-ai-in-advertising-striking-the-balance-between-creativity-and-responsibility.html) that an AI image showing cosmetic results the product doesn’t achieve is likely to mislead, and that a note saying AI was used won’t fix it. **Rights.** Read the provider’s terms. [Runway’s terms](https://runway.com/terms-of-use) say it doesn’t claim ownership of your outputs or restrict their commercial use, and that it may use your inputs and outputs to train its models. Owning an output isn’t the same as holding copyright in it: the US Copyright Office [concluded in January 2025](https://www.copyright.gov/newsnet/2025/1060.html) that AI output is protected only where a human determined enough of its expressive elements, and that prompts alone are not enough. Leave out real people’s faces and voices, other companies’ brands, and music you don’t have rights to. **People in the ad.** In the US, the FTC says its rule on reviews and testimonials has [no blanket ban on AI avatars](https://www.ftc.gov/business-guidance/resources/consumer-reviews-testimonials-rule-questions-answers). A fake testimonial is banned, though, and so is a celebrity avatar used without permission if people would think the celebrity endorsed you. ## Which AI videos need a label? - **TikTok ads.** TikTok [requires an AI-generated content disclaimer](https://ads.tiktok.com/help/article/about-ad-disclaimers-in-tiktok-ads-manager) on ads that are completely AI-generated or significantly modified by AI. - **YouTube.** Creators [must disclose](https://support.google.com/youtube/answer/14328491) realistic AI content, including a realistic scene that didn’t actually happen, and YouTube may add a label itself when a video carries C2PA metadata ([Content Credentials](https://c2pa.org/), an open standard for showing where content came from). - **Meta.** Ads made with Meta’s own AI tools, or detected as made with other AI tools, get an “AI info” label. - **The EU.** Deepfakes, meaning content that would falsely appear real, need a clear label, and providers must mark AI content in a machine-readable way. Veo videos, for example, carry Google’s SynthID watermark. - **The UK.** The ASA says there is no blanket legal requirement to disclose AI in ads; the test is whether the ad would mislead without a disclosure. This is a summary, not legal advice. ## What does a good batch look like? Here is the kind of request you might give your AI. Illustrative: > “Turn this product page into 20 vertical 15-second ads, each with a different hook. Tell me what it will cost before you start, and give me a sheet with each take’s prompt, model and the label it needs.” A good result quotes the cost first, uses one model so the takes match, returns each file with its prompt and the model that made it, and flags every take that shows realistic people or scenes. Comparing plans? See [pay per call or a monthly plan](https://ikag.dev/blog/pay-per-call-vs-monthly-plans-for-ai-tools). ## Where IKAG fits AI video in bulk is one of the [next kinds of tools](https://ikag.dev/#what) IKAG is building: your AI would ask for the takes, IKAG would quote the call before it runs, and you would pay per call instead of a subscription to a video app. No prices have been set. IKAG is opening to early users and data partners first. ## Questions ### Can I make AI videos without a subscription? Yes. Google prices its Veo 3.1 model in the Gemini API per second of generated video, usually paid from prepaid credits that expire after 12 months, and Runway sells credits for its developer API. You need a developer account and a key, or a tool that calls the model for you. ### Do AI-generated ads have to be labelled? Often. TikTok requires an AI-generated content disclaimer on ads that are completely AI-generated or significantly modified by AI, Meta labels ads it detects as made with AI tools, and EU rules require deepfakes to be labelled from 2 August 2026. The UK has no blanket legal requirement, but ads must not mislead. This is not legal advice. ### Who owns an AI-generated video? Read the provider’s terms: Runway’s, for example, say it does not claim ownership of outputs or restrict their commercial use. Copyright is a separate question. The US Copyright Office says prompts alone do not make you the author. ### What happened to Sora? OpenAI discontinued the Sora app and website on 26 April 2026, and removed its Videos API and Sora 2 models from its API on 24 September 2026, with no replacement listed. ### Can I use AI avatars in ads? In the US, the FTC says its rule on reviews and testimonials has no blanket ban on AI avatars. A fake or false testimonial is banned, though, and so is a celebrity avatar used without permission if people would think the celebrity endorsed the product. ## Sources - [Runway: Pricing (checked 17 October 2026)](https://runway.com/pricing) - [Runway: API pricing and costs (checked 17 October 2026)](https://docs.dev.runwayml.com/guides/pricing/) - [Runway: Terms of Use (last updated 15 September 2026)](https://runway.com/terms-of-use) - [Google: Google AI Plus, Pro and Ultra plans (checked 17 October 2026)](https://gemini.google/subscriptions/) - [Google AI for Developers: Gemini Developer API pricing (last updated 16 October 2026)](https://ai.google.dev/gemini-api/docs/pricing) - [Google AI for Developers: Gemini API billing (last updated 28 September 2026)](https://ai.google.dev/gemini-api/docs/billing) - [Google AI for Developers: Generate videos with Veo 3.1 in the Gemini API (last updated 8 October 2026)](https://ai.google.dev/gemini-api/docs/veo) - [OpenAI Help Center: What to know about the Sora discontinuation (checked 17 October 2026)](https://help.openai.com/en/articles/20001152-what-to-know-about-the-sora-discontinuation) - [OpenAI API: Deprecations (checked 17 October 2026)](https://developers.openai.com/api/docs/deprecations) - [Meta: Expanding GenAI transparency for Meta’s ads products (3 February 2025, updated 1 June 2026)](https://about.fb.com/news/2025/02/gen-ai-transparency-metas-ads-products/) - [European Commission: Quick facts, transparency rules for AI systems (last update 29 July 2026)](https://digital-strategy.ec.europa.eu/en/factpages/quick-facts-transparency-rules-ai-systems) - [TikTok for Business: About ad disclaimers in TikTok Ads Manager (last updated September 2025)](https://ads.tiktok.com/help/article/about-ad-disclaimers-in-tiktok-ads-manager) - [YouTube Help: Disclosing use of GenAI content (checked 17 October 2026)](https://support.google.com/youtube/answer/14328491) - [C2PA: Content Credentials (checked 17 October 2026)](https://c2pa.org/) - [ASA and CAP: Disclosure of AI in advertising (29 May 2025)](https://www.asa.org.uk/news/disclosure-of-ai-in-advertising-striking-the-balance-between-creativity-and-responsibility.html) - [US Copyright Office: Copyright Office releases Part 2 of Artificial Intelligence Report (29 January 2025)](https://www.copyright.gov/newsnet/2025/1060.html) - [FTC: The Consumer Reviews and Testimonials Rule, questions and answers (November 2024)](https://www.ftc.gov/business-guidance/resources/consumer-reviews-testimonials-rule-questions-answers) --- # ChatGPT Dots, Grok Bot and Meta Muse explained > ChatGPT Dots, Grok Bot and Meta’s Muse are AI agents with their own cloud computers. What each one does, who can get it and what none of them can reach. - URL: https://ikag.dev/blog/chatgpt-dots-grok-bot-meta-muse-explained - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-17 - Updated: 2026-10-17 - Reading time: 6 min - Publisher: IKAG ## Overview ChatGPT Dots, Grok Bot and Meta’s Muse are always-on AI agents announced between August and September 2026. Each works on a computer of its own in the cloud, signs in to the apps you connect and keeps going while you do something else. They differ in where you talk to them and which plans include them. None of the three announcements offers paid data of its own: the agents work through your logins, your connected apps and the open web. ## Key points - ChatGPT Dots, Grok Bot and Meta’s Muse are AI agents that run tasks on a computer of their own in the cloud, announced between 11 August and 29 September 2026. - OpenAI includes ChatGPT Dots with Pro outside the EEA, Switzerland and the UK, and with Business Premium; Grok Bot comes with Premium, Super or Ultra on The X Subscription and with legacy Cursor plans; Muse is free for most uses in the US and Canada. - All three act through the apps you connect and the websites their browsers can open, and check with you before sensitive actions such as sending an email or buying something. - None of the three announcements offers paid data of its own, so a lead database, a listings service or a video model still needs its own account, or a connected tool that pays for it. OpenAI’s ChatGPT Dots, Grok Bot and Meta’s Muse are AI agents you hand work to, not just chat with. Each has a computer of its own in the cloud, signs in to the apps you connect and keeps working after you close the app, checking with you before anything sensitive. They arrived within seven weeks of each other: Grok Bot on 11 August 2026, Muse on 8 September and ChatGPT Dots on 29 September. What they share matters as much as how they differ. None of the three announcements offers paid data of its own. The agents work through your logins, so what they can reach is mostly what you already have. ## What are ChatGPT Dots? OpenAI [introduced ChatGPT Dots](https://openai.com/index/introducing-dots/) on 29 September 2026 as always-on agents that live in ChatGPT. Each one runs on GPT-6 Astra, has its own cloud computer and browser, and can connect to over 4,000 apps through ChatGPT’s plugins. You can message or call your agent in ChatGPT on desktop, web and mobile, and message it in Slack and Teams. Each agent can research in the background, but OpenAI restricts those background tools to reading: they can’t send messages, change app content or control a browser. Everything else follows rules you can adjust, and some steps, such as changing a password, always stay with you. According to [OpenAI’s help centre](https://help.openai.com/en/articles/20001530-getting-started-with-your-dot), ChatGPT Dots are rolling out to Pro users aged 18 and over outside the European Economic Area, Switzerland and the UK, and to Business Premium users in every supported ChatGPT region. Enterprise, Edu and Healthcare workspaces get a beta that an admin has to turn on. Your first agent is included in Pro or Business Premium at no extra cost. ## What is Grok Bot? [Grok Bot](https://x.ai/news/introducing-grok-bot) was announced as a beta on 11 August 2026: a team of always-on agents that share a computer of their own in the cloud. Bots sign in to the tools you already use and work across apps and websites, including, in the announcement’s words, platforms with “no clean API or MCP”. You message a bot from your phone or desktop like a colleague, put several in a group chat to divide the work, or show one a task once so it can repeat it as a routine. The apps are for desktop and iOS. Grok Bot’s [plans and billing page](https://cursor.com/help/grok-bot/plans), hosted by Cursor, says it is included with Premium, Super and Ultra on The X Subscription, but not Plus or Free, and with legacy paid Cursor plans and Cursor Teams. Linking a SuperGrok, SuperGrok Plus, SuperGrok Heavy or X Premium+ subscription also gives access, as does a seat on a self-serve Grok Business plan. ## What is Meta’s Muse? Meta [introduced Muse](https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/) on 8 September 2026 as a personal agent that does the work rather than only answering questions. It runs on Muse Secure VM, a dedicated cloud computer with its own browser, and you talk to it in the Muse app or in WhatsApp. Meta’s examples include sending email, booking travel, filling in forms and negotiating on your behalf, and Muse can pay at checkout with Link by Stripe. It asks for approval before sending an email or making a purchase. Meta says Muse is free for most of what people need, with subscription plans for people who want more. It started in the US on iOS, Android and muse.ai, and Meta’s [small business announcement](https://about.fb.com/news/2026/09/introducing-muse-small-business/) of 29 September describes it as available in the US and Canada. ## How do ChatGPT Dots, Grok Bot and Muse compare? | At a glance | ChatGPT Dots (OpenAI) | Grok Bot | Muse (Meta) | | --------------------- | --------------------------------------------------------------------------- | ------------------------------------------------------------------------------------ | -------------------------------------------- | | Announced | 29 September 2026 | 11 August 2026, as a beta | 8 September 2026 | | Works on | Its own cloud computer and browser | A cloud computer its bots share | Muse Secure VM, with its own browser | | Where you talk to it | ChatGPT, Slack, Teams | Desktop and iOS apps | Muse app, muse.ai, WhatsApp | | Who can get it | ChatGPT Pro (not EEA, Switzerland or UK), Business Premium; Enterprise beta | Premium, Super or Ultra on The X Subscription; legacy Cursor plans; linked SuperGrok | Free for most uses, in the US and Canada | | Adding your own tools | ChatGPT plugins, including custom MCP servers on some plans | Ask in chat to add an MCP server | Custom connectors built from a service’s API | | Paid data of its own | None described | None described | None described | ## What can’t they reach on their own? All three are built to act through accounts you already have. ChatGPT Dots use the apps you connect and their browsers, Grok Bot signs in to your tools, and Muse uses connectors and the services you sign it in to. None of the three announcements describes a paid data source of its own. That matters when a task needs something behind a paywall. Lists of businesses and decision makers with their work contact details, AI video in bulk and detailed local business listings often come from paid services, each with its own subscription. For the agent to use one, you sign it in to an account that pays for it. Even where a ready-made connector exists, the subscription stays: [Grok’s connector catalogue](https://docs.x.ai/grok/connectors) notes that its Meltwater connector requires an active Meltwater subscription. We look at why those subscriptions add up, even when one account works across several AI apps, in [whether you need a subscription for every AI tool](https://ikag.dev/blog/do-you-need-a-subscription-for-every-ai-tool). Websites can also say no. In September 2026 Amazon blocked Muse from shopping on Amazon.com, and told [Adweek](https://www.adweek.com/commerce/amazon-locks-out-metas-muse-in-agentic-shopping-standoff/) it had not been made aware in advance that Muse would access its store. An agent that works through a browser depends on each site’s rules. ## Can you add your own tools to them? Yes, in different ways, and two of them use MCP, an open standard for connecting AI apps to outside tools, which we explain in [what is MCP](https://ikag.dev/blog/what-is-mcp-plain-english-guide). - **ChatGPT Dots** use the plugins you set up in ChatGPT: OpenAI says [plugin permissions are shared](https://help.openai.com/en/articles/20001529-dots-privacy-security-and-safety-faqs) across ChatGPT Dots, ChatGPT, ChatGPT Work and Codex. ChatGPT adds a custom MCP server as a plugin on the web, [in full on Business, Enterprise and Edu](https://help.openai.com/en/articles/12584461-connecting-custom-mcp-servers-in-chatgpt) and with read and fetch permissions only on Pro. OpenAI notes that the apps available to the agents may differ. - **Grok Bot** has no settings form for this. A Cursor support engineer [explained on 30 August 2026](https://forum.cursor.com/t/grokbot-custom-connectors/169965) that you ask your bot in chat to add an MCP server, giving its address, and it confirms the details before setting it up. The server must be reachable over the internet, because the bot runs in the cloud. That is a forum answer, not official documentation. - **Muse** has built-in connectors under Settings, and you can ask it to build a custom connector for another service, which can involve retrieving that service’s API details. [Meta’s help page](https://www.meta.com/help/artificial-intelligence/1687253048996149/) doesn’t mention MCP, and Meta doesn’t review custom connectors. The steps for ChatGPT, Claude, Gemini and coding tools are in [how to add a tool to ChatGPT, Claude and Gemini](https://ikag.dev/blog/how-to-add-a-tool-to-chatgpt-claude-gemini). ## Which one should you choose? Start from where you already work and pay. If your day runs through ChatGPT, Slack or Teams and you have Business Premium, or Pro outside the EEA, Switzerland and the UK, ChatGPT Dots are already included. If you pay for Premium or above on The X Subscription, or for a paid Cursor plan, Grok Bot is likely included too. If you want help with errands and shopping in the US or Canada without a paid plan, Muse is free for most uses. Whichever you pick, look at what it is signed in to: that is the edge of what it can do. ## Where IKAG fits IKAG is building one connection that gives the AI people already use, agents included, the paid tools and data they can’t reach on their own: lead data, AI video, web data, local business listings, social media data and company records. The direction is pay per call instead of a subscription to each service, and no prices have been set. In most apps, connecting is designed to take one link and a sign-in. IKAG is opening to early users and data partners first. ## Questions ### What are ChatGPT Dots? ChatGPT Dots are OpenAI’s always-on agents in ChatGPT, announced on 29 September 2026. Each one runs on GPT-6 Astra with its own cloud computer and browser, uses the apps you connect through ChatGPT’s plugins, and can be reached in ChatGPT, Slack and Teams. ### Who can use Grok Bot? Grok Bot’s plans page says it is included with Premium, Super and Ultra on The X Subscription, and with legacy paid Cursor plans and Cursor Teams. Linking a SuperGrok, SuperGrok Plus, SuperGrok Heavy or X Premium+ subscription also gives access, as does a seat on a self-serve Grok Business plan. ### Is Meta Muse free? Meta says Muse is free for most of what people need, with subscription plans for people who want to do more. Meta describes it as available in the US and Canada, in the Muse app, on muse.ai and in WhatsApp. ### Can ChatGPT Dots, Grok Bot or Muse use paid data? Only through accounts and tools you connect. None of the three announcements offers paid data of its own, so a paid database or service needs its own subscription, or a connected tool that pays for it. ## Sources - [OpenAI: Introducing dots (29 September 2026)](https://openai.com/index/introducing-dots/) - [OpenAI Help Center: Getting started with your dot (updated October 2026)](https://help.openai.com/en/articles/20001530-getting-started-with-your-dot) - [OpenAI Help Center: Dots privacy, security, and safety FAQs (updated October 2026)](https://help.openai.com/en/articles/20001529-dots-privacy-security-and-safety-faqs) - [OpenAI Help Center: Connecting custom MCP servers in ChatGPT (updated October 2026)](https://help.openai.com/en/articles/12584461-connecting-custom-mcp-servers-in-chatgpt) - [x.ai: Introducing Grok Bot (11 August 2026)](https://x.ai/news/introducing-grok-bot) - [Cursor Docs: Grok Bot plans and billing (checked 17 October 2026)](https://cursor.com/help/grok-bot/plans) - [Cursor Community Forum: Grokbot custom connectors, answer from Cursor support (30 August 2026)](https://forum.cursor.com/t/grokbot-custom-connectors/169965) - [Grok docs: Connectors (updated 10 October 2026)](https://docs.x.ai/grok/connectors) - [Meta: Introducing Muse (8 September 2026)](https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/) - [Meta: Muse for Small Business (29 September 2026)](https://about.fb.com/news/2026/09/introducing-muse-small-business/) - [Meta Help Centre: How Muse works with Connectors (checked 17 October 2026)](https://www.meta.com/help/artificial-intelligence/1687253048996149/) - [Adweek: Amazon locks out Meta’s Muse in agentic shopping standoff (21 September 2026)](https://www.adweek.com/commerce/amazon-locks-out-metas-muse-in-agentic-shopping-standoff/) --- # Do you need a subscription for every AI tool? > Not one per AI app: a single account at a service can work in ChatGPT, Claude and Cursor. What adds up is a subscription to each service your AI uses. - URL: https://ikag.dev/blog/do-you-need-a-subscription-for-every-ai-tool - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-17 - Updated: 2026-10-17 - Reading time: 7 min - Publisher: IKAG ## Overview No AI app needs its own subscription to a service. Many services publish an MCP server or their own connectors, so one account works in ChatGPT, Claude, Cursor and other AI apps: Apollo.io, for one, connects a single Apollo account to all three, and to Perplexity and Replit, at no extra cost. What adds up is the services. Lead data, AI video, web data and social listening usually come from different companies, each sold as a subscription, and your AI may use each one only a few times a month. ## Key points - No AI app needs its own subscription to a service: through MCP or the service’s own connectors, one account can work in ChatGPT, Claude, Cursor and many other AI apps. - Apollo.io, a sales platform, lists built-in connectors for ChatGPT, Claude, Perplexity, Replit and Cursor at no extra cost, and your Apollo plan, permissions and credits apply in each. - MCP fixes the plug, not the subscriptions: lead data, AI video, web data and social listening usually come from different companies, each sold as a subscription, and Zylo found organisations leave 36% of their software licences unused on average. - ChatGPT Dots, Grok Bot and Meta’s Muse work through the accounts you connect, and none of their announcements offers paid data of its own. - One connection that holds the provider agreements and charges per call is the alternative to a subscription for each service. IKAG is building one, and no prices have been set. IKAG is opening to early users and data partners first. No AI app needs its own subscription to a service. One account at a service, such as a lead database, can be connected to ChatGPT, Claude, Cursor and many other AI apps through MCP or the service’s own connectors, and each of them uses that same account and plan. But each service your AI uses is still its own subscription. Lead data, AI video, web data and social listening usually come from different companies, each with a monthly plan, and your AI may use each one only a few times a month. That is what adds up. Here is why it happens, what it costs and what would change it. ## Can one subscription work in ChatGPT, Claude and Cursor? Yes, for many services. The usual route is MCP, the Model Context Protocol, an open standard for connecting AI apps to outside tools. A service publishes one MCP server, and AI apps that support MCP can connect to it with your account. The standard spread quickly: by December 2025 Anthropic counted [more than 10,000 active public MCP servers](https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation), and ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code had adopted it. Apollo.io, a sales platform and lead database, shows how this works. Its [MCP documentation](https://docs.apollo.io/docs/apollo-mcp) lists built-in connectors for ChatGPT, Claude, Perplexity, Replit and Cursor, and a standalone MCP server for other apps that support one. Each connects to your Apollo account. Apollo says there is no additional cost to connect, and that your Apollo plan, permissions and credits still apply: one subscription, used from any of those apps. Connecting has got simpler, too. MCP’s specification lets a connector [ask you to sign in](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization) with OAuth, the kind of sign-in and permission screen many apps use, instead of making you paste a key, and Apollo’s connectors work this way. [How to add a tool to ChatGPT, Claude and Gemini](https://ikag.dev/blog/how-to-add-a-tool-to-chatgpt-claude-gemini) has the steps for each app. ## Does MCP get rid of the subscriptions? No. MCP fixes the plug, not the subscriptions. A connector to a paid service is still a connector to that service’s account, plan and terms, and many of the tools your AI might need are paid services: lead data (businesses and decision makers, with their work contact details), AI video, web data, social listening and local business listings, each usually from a different company. **Illustrative example.** Four requests in one week, needing lead data, AI video, web data and social media data, can mean four sign-ups and four monthly plans before your AI makes a single call. Connecting each service to three AI apps doesn’t change the count: it’s still four subscriptions, and each one renews whether your AI used it twice or not at all. [MCP in plain English](https://ikag.dev/blog/what-is-mcp-plain-english-guide) explains how the plug works. ## Why do services sell access through accounts and plans? Because selling access to a tool or a dataset means knowing who is using it. For the service, an account does four jobs: - **It identifies the caller.** Each request arrives with a sign-in token or a key that says which account sent it. - **It limits abuse.** New accounts often start with low limits. Anthropic, for example, says the lower [starting limits](https://platform.claude.com/docs/en/api/rate-limits) on its API are part of how it prevents fraud and abuse. - **It carries the licence terms.** Data providers license their data for certain uses, and the account is where you accepted them. Apollo, for one, [prohibits AI model training](https://docs.apollo.io/docs/apollo-mcp) through its MCP connections. - **It gets the service paid.** A card on file, and often a monthly plan, which brings in the same revenue whether you use the service or not. None of that is unreasonable for one service. The trouble starts when your AI needs a different one for each kind of request, and uses each only now and then. ## What do all those subscriptions cost? **Money.** A plan charges for the month whether your AI calls the service once or a thousand times. [Zylo’s 2026 index](https://zylo.com/news/2026-saas-management-index) found that organisations leave 36% of their software licences unused on average, and that ChatGPT is now the most expensed app, a sign of AI tools bought on individual cards, outside the usual buying process. [Subscription fatigue](https://ikag.dev/blog/subscription-fatigue-tools-you-use-twice) covers what that looks like for companies and households, and [pay per call vs monthly plans](https://ikag.dev/blog/pay-per-call-vs-monthly-plans-for-ai-tools) compares the two ways of paying. **Time.** Each service means a sign-up, a plan to compare, a card to enter and a renewal date to remember. For one service it’s a chore. For six, it’s a project. **Attention.** Each service has its own dashboard, usage page and invoice. Working out what your AI spent last month means visiting all of them. ## What about API keys? Not every connection uses sign-in. Some services and some AI apps still connect with an API key: a secret code that identifies your account to a service’s API, the door software uses to talk to it. When your AI calls the service, it sends the key, and the service knows whom to serve and whom to bill. So every key is a password your AI holds for you, and keys leak. GitGuardian counted [28.65 million new secrets](https://blog.gitguardian.com/the-state-of-secrets-sprawl-2026/), such as API keys, added to public GitHub code in 2025, 34% more than the year before. Secrets for AI services rose 81%, and it found 24,008 unique secrets in MCP configuration files on public GitHub, the files that tell AI apps how to reach their tools. A key pasted into three AI apps is three copies to look after, so prefer sign-in where a service offers it. ## Do ChatGPT Dots, Grok Bot and Muse change this? Not on their own. The newest agents can do a lot of work by themselves, but they do it through the accounts you connect. - **ChatGPT Dots** (OpenAI, 29 September 2026): each has its own cloud computer and browser, and works through [the apps you’ve connected](https://openai.com/index/introducing-dots/), with over 4,000 apps available through plugins. - **Grok Bot** (11 August 2026): the bots have their own computer and [sign in to the tools you already use](https://x.ai/news/introducing-grok-bot). - **Muse** (Meta, 8 September 2026): Muse runs on its own secure virtual machine and works across [the apps a person connects](https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/), keeping their credentials in secure storage. None of the three announcements offers paid data of its own. To use a paid lead database, a video tool or a scraper, an agent still needs an account with each one, and that means yours. Muse can even pay at checkout with Stripe’s Link, but paying for a plan on your behalf still leaves you with the plan. [Our guide to ChatGPT Dots, Grok Bot and Muse](https://ikag.dev/blog/chatgpt-dots-grok-bot-meta-muse-explained) goes further. ## How does one connection that charges per call change it? Put the subscriptions in one place. Instead of you keeping a plan with each service, your AI connects once to a service that holds agreements with many providers, calls the right one for each request and charges per call. You pay for the calls your AI makes, not for months of access it barely used, and like any MCP server, that one connection can be added to several AI apps. A plan can still be the better deal for a service your AI uses every day. Some services already work this way for certain kinds of tools, with different trade-offs, which we compare in [tool routers and data networks](https://ikag.dev/blog/tool-routers-and-data-networks-for-ai-agents). | Question | A subscription to each service | One connection, per call (IKAG’s design) | | --------------------------- | ---------------------------------- | -------------------------------------------------- | | AI apps it works in | Many, with one account per service | Many, with one connection | | Accounts to open | One per service | One, for the services it manages | | How you pay | A plan per service, used or not | Per call, quoted before it runs | | Keys and sign-ins | One per service | Sign-in in most apps, a key where an app needs one | | Where spending shows up | Each service’s dashboard | Each call itemised on a receipt | | A service you use every day | Its plan may be the better deal | Keep that plan and use your own key with it | The last column is a design, not a price list. Managed tools depend on an agreement with each provider, and IKAG names none until one is agreed; Apollo.io appears above only as an example, not as an IKAG provider. No prices have been set. With your own provider account, that provider’s plan and charges still apply, and IKAG may charge separately for its part. ## Where IKAG fits IKAG is building that one connection, designed so the AI you already use can call paid tools such as lead data, AI video, web data and social media data, and pay per call instead of you keeping a subscription to each service. No prices have been set. Its first tool, company records from official registries, already runs inside IKAG, and in most apps connecting is designed to take one link and a sign-in. Data providers can see how a pilot works on the [partners page](https://ikag.dev/partners). IKAG is opening to early users and data partners first. ## Questions ### Do I need a separate subscription to a service for each AI app? No. Many services publish an MCP server or their own connectors, so one account can work in ChatGPT, Claude, Cursor and other AI apps. Apollo.io, for example, lists built-in connectors for ChatGPT, Claude, Perplexity, Replit and Cursor at no extra cost, and your Apollo plan and credits apply in each. ### Why do the subscriptions still add up? Because each service your AI uses sells access on its own terms. Lead data, AI video, web data and social listening usually come from different companies, each with its own account and plan, and your AI might use each one only a few times a month. ### Does MCP get rid of subscriptions? No. MCP standardises how AI apps connect to services, so one account can work in many apps, and its specification supports signing in with OAuth instead of pasting a key. Each paid service still decides who may use it and on what plan. ### Do ChatGPT Dots, Grok Bot or Meta’s Muse come with paid data? Their announcements describe agents that work through the apps and accounts you connect. None of them offers paid data of its own, so paid services still depend on subscriptions you hold. ### Is it safe to paste an API key into an AI app? Treat every key like a password: share it only with apps you trust, never paste it into a chat or a public file, and remove it when you stop using the service. Where a service offers sign-in instead of a key, prefer sign-in. ### How would one connection that charges per call change this? A single connection could hold the agreements with providers, so your AI could reach many services through it and pay per call for the ones it manages, instead of you keeping a subscription to each. IKAG is building this. No prices have been set. IKAG is opening to early users and data partners first. ## Sources - [Apollo.io docs: Apollo MCP (updated 5 October 2026)](https://docs.apollo.io/docs/apollo-mcp) - [Anthropic: Donating the Model Context Protocol and establishing the Agentic AI Foundation (9 December 2025)](https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation) - [Model Context Protocol specification, version 2026-07-28: Authorization](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization) - [Anthropic, Claude Platform docs: Rate limits and spend limits (accessed 17 October 2026)](https://platform.claude.com/docs/en/api/rate-limits) - [Zylo: 2026 SaaS Management Index finds AI-native app adoption is surging (29 January 2026)](https://zylo.com/news/2026-saas-management-index) - [GitGuardian: The State of Secrets Sprawl 2026 (17 March 2026)](https://blog.gitguardian.com/the-state-of-secrets-sprawl-2026/) - [OpenAI: Introducing dots (29 September 2026)](https://openai.com/index/introducing-dots/) - [x.ai: Introducing Grok Bot (11 August 2026)](https://x.ai/news/introducing-grok-bot) - [Meta: Introducing Muse: The World’s First Personal AI Agent Built for Everyone (8 September 2026)](https://about.fb.com/news/2026/09/introducing-muse-personal-ai-agent/) --- # How AI agents find B2B leads, and what to check > Where an AI agent’s list of businesses and decision makers comes from, why models alone get work emails wrong, the UK and US email rules, and what a good list includes. - URL: https://ikag.dev/blog/how-ai-agents-find-b2b-leads - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-17 - Updated: 2026-10-17 - Reading time: 5 min - Publisher: IKAG ## Overview An AI agent finds B2B leads by looking them up, not by remembering them: company registries say who a business legally is, its own website says who does what, and B2B contact databases hold work email addresses. Without those tools a model guesses, and guesses badly: in a September 2026 test by Hunter, 76% of the checkable email addresses that models produced were invalid. A good list names the source and date of every detail, lists what it couldn’t find, and follows the email rules where you send. ## Key points - An AI model doesn’t hold a directory of work email addresses. To build a lead list it needs tools that reach company registries, business websites and B2B contact databases. - In a test Hunter published in September 2026, 76% of the email addresses that language models produced on their own, and that its verifier could check, were invalid. Hunter sells email-finding tools. - Company registries confirm a business’s legal name, status and officers, but not its inbox: Companies House doesn’t publish the email address each limited company gives it. - In the UK, PECR lets you email limited companies without consent if you say who you are and offer an opt-out, but a named person’s work email is personal data under UK GDPR. In the US, CAN-SPAM makes no exception for business email. - A good lead list gives the source and date of every detail, says which email addresses were verified, and lists what it could not find. An AI agent can build a list of businesses and decision makers, with their work contact details, but not from memory. A language model doesn’t hold a directory of who works where, and when it is pushed for an email address it tends to make one up. The agent has to look each detail up in a source that holds it, and a good list tells you which source that was. ## Where does B2B lead data come from? Most lead lists combine three kinds of source. **Company registries.** A registry records who a business legally is. In the UK, Companies House gives anyone a company’s [registered address, date of incorporation and current and resigned officers](https://www.gov.uk/get-information-about-a-company) at no charge, and its [register](https://find-and-update.company-information.service.gov.uk/) shows whether the company is active, with the note that Companies House doesn’t check the accuracy of what is filed. What a registry doesn’t give you is an inbox. A limited company must give Companies House an email address when it is set up, but that address is [not published on the public register](https://www.gov.uk/limited-company-formation/company-address). **The businesses’ own websites.** Team pages, contact pages and press releases say who does what, and often list a work email or a general inbox. Reading them is ordinary web data work, with the usual rules about each site’s terms and robots.txt (see [how AI agents get live data](https://ikag.dev/blog/how-ai-agents-get-live-data)). **B2B contact databases.** Commercial databases hold named contacts and work email addresses at scale. ZoomInfo, for example, [says it builds its database](https://www.zoominfo.com/data-sources) from public information on more than 28 million sites and domains a day, data partners, an in-house research team, and more than 200,000 professionals who share business contacts in exchange for access to its tool. A database like this is broad, but you are relying on its sourcing, and people change jobs. | Source | What it gives | What it lacks | Watch for | | ---------------------- | ---------------------------------------- | ----------------------------- | ------------------------------- | | Company registry | Legal name, number, status, officers | Email addresses | Trading and legal names differ | | The business’s website | Who does what, published contact details | A date on most pages | Site terms; pages that go stale | | B2B contact database | Named contacts and work emails at scale | How each record was collected | Job changes; your lawful basis | | The model on its own | Plausible guesses | Any check at all | Invented addresses | ## Why can’t the AI just know the email address? Because it is guessing. In a [test Hunter published in September 2026](https://hunter.io/blog/llms-get-email-data-wrong/), language models were asked for the work email addresses of people in roles such as head of marketing at 125 real companies, and every address they returned went through Hunter’s email verifier. Of the 1,604 addresses that got a clear verdict, 76% were invalid, 10% could not be confirmed either way and 14% were valid. When prompts pushed for a best guess, models filled the gaps with addresses they had made up. Hunter sells email-finding and verification tools, so read its numbers with that in mind. Its own comparison makes the point, though: the strongest model it tested ended with a valid address for 19.5% of requests on its own, and for 78.1% when it could call Hunter’s lookup tool. A model is good at deciding who to contact and why. The address has to come from a source that can check it. ## What are the rules for emailing businesses? Who you may email, and how, depends on where you and they are. In outline: - **UK, PECR.** The [ICO’s guidance on business-to-business marketing](https://ico.org.uk/for-organisations/direct-marketing-and-privacy-and-electronic-communications/business-to-business-marketing/) says the PECR rule on marketing by email doesn’t apply to “corporate subscribers” such as limited companies and limited liability partnerships, so you don’t need their consent. You must say who you are and give a valid address for opting out. Sole traders and some partnerships count as individuals, so you need their consent or the “soft opt-in”. If you can’t tell which a contact is, the ICO says to treat them as an individual. - **UK, UK GDPR.** A work email that identifies a person, such as a first name and surname at a company, is personal data. You need a lawful basis, often legitimate interests after a three-part test, and if you got the data from somewhere other than the person, you must usually tell them within a month. People have an absolute right to object to direct marketing. The ICO says this guidance is under review after the [Data (Use and Access) Act 2025](https://www.legislation.gov.uk/ukpga/2025/18/contents). - **US, CAN-SPAM.** The [FTC’s compliance guide](https://www.ftc.gov/business-guidance/resources/can-spam-act-compliance-guide-business) says the law “makes no exception for business-to-business email”. Commercial messages need honest headers and subject lines, a valid postal address and a clear way to opt out, and opt-outs must be honoured within 10 business days. Hiring someone else to send them doesn’t move the responsibility. - **Platform terms.** Sites have rules of their own. [LinkedIn’s User Agreement](https://www.linkedin.com/legal/user-agreement), for example, forbids software that scrapes or copies its services, including profiles. The ICO also notes that messaging someone on a professional networking site is often not business-to-business marketing at all, because people use those sites in a personal, professional capacity. This is a summary, not legal advice. Check the rules for your own campaign, and ask a lawyer when the stakes are high. ## What should a good lead list include? Here is the kind of request you might give your AI. Illustrative: > “Find 40 dental practices in Bristol and the practice manager at each, with their work contact details. Say where each detail came from and when, and flag any practice that’s a sole trader or partnership.” A list you can act on has, for each row: - **The business:** its legal name and registry number, because names aren’t unique, and whether the registry shows it as active. - **Its legal form:** a company, or a sole trader or partnership, because PECR treats them differently. - **The decision maker:** their role, and the page that names them. - **The work email address:** with its verification result (valid, can’t be confirmed, or invalid) and the date it was checked. - **Sources and dates:** where every value came from and when, as described in [what the source record behind an AI answer contains](https://ikag.dev/blog/what-the-source-record-behind-an-ai-answer-contains). - **Gaps, listed:** “no named manager on the practice’s website” is more useful than a guessed name. - **Opt-outs:** checked against your own suppression list before anything is sent. Fewer rows with sources beat a full sheet of guesses, and the [five-step check for AI answers](https://ikag.dev/blog/how-to-check-ai-answers) works on lead lists too. ## Where IKAG fits Lead data is one of the [next kinds of tools](https://ikag.dev/#what) IKAG is building: your AI would ask for businesses and decision makers, with their work contact details, and IKAG is designed to return each detail with where it came from, where the source supports it. You would pay per call instead of a subscription to a lead database, and no prices have been set. IKAG’s first tool, company records from GLEIF, SEC EDGAR and Wikidata, already runs inside it. IKAG is opening to early users and data partners first. ## Questions ### Can an AI assistant find someone’s work email address? Only if it can reach a source that holds it, such as the company’s own website or a B2B contact database. On their own, models tend to guess: in Hunter’s September 2026 test, 76% of the checkable addresses that models produced were invalid. ### Do I need consent to email businesses in the UK? Not for limited companies and other corporate bodies under PECR, as long as you say who you are and give a valid address to opt out. Sole traders and some partnerships need consent or the soft opt-in, and a named person’s work email is still personal data under UK GDPR. This is not legal advice. ### Does CAN-SPAM apply to business-to-business email? Yes. The FTC says the law makes no exception for business-to-business email. Commercial messages need honest headers and subject lines, a valid postal address and a clear way to opt out, and opt-outs must be honoured within 10 business days. ### Can an AI agent collect contacts from professional networking sites? Check the site’s terms first: LinkedIn’s User Agreement, for example, forbids software that scrapes or copies its services, including profiles. The ICO adds that people on such sites are often there in a personal, professional capacity, so UK GDPR and PECR still apply. ### What should a B2B lead list include? For each business: its legal name and registry number, whether it is active, the decision maker’s role, a work email address with its verification result, where each detail came from and when, and a note of anything that could not be found. ## Sources - [Hunter: LLMs get 76% of emails wrong, but Hunter can easily fix it (1 September 2026)](https://hunter.io/blog/llms-get-email-data-wrong/) - [ICO: Business-to-business marketing (guidance under review after the Data (Use and Access) Act; checked 17 October 2026)](https://ico.org.uk/for-organisations/direct-marketing-and-privacy-and-electronic-communications/business-to-business-marketing/) - [legislation.gov.uk: Data (Use and Access) Act 2025](https://www.legislation.gov.uk/ukpga/2025/18/contents) - [FTC: CAN-SPAM Act, a compliance guide for business (August 2023)](https://www.ftc.gov/business-guidance/resources/can-spam-act-compliance-guide-business) - [LinkedIn: User Agreement (effective 3 November 2025)](https://www.linkedin.com/legal/user-agreement) - [GOV.UK: Get information about a company (checked 17 October 2026)](https://www.gov.uk/get-information-about-a-company) - [Companies House: Find and update company information (checked 17 October 2026)](https://find-and-update.company-information.service.gov.uk/) - [GOV.UK: Check the rules for registered office addresses and email addresses (checked 17 October 2026)](https://www.gov.uk/limited-company-formation/company-address) - [ZoomInfo: Data sources, where we get our data (checked 17 October 2026)](https://www.zoominfo.com/data-sources) --- # How to add a tool to ChatGPT, Claude and Gemini > Step by step: add a custom connector or MCP server in ChatGPT, Claude, Gemini, Cursor, VS Code, Claude Code and Codex, with plan limits as of October 2026. - URL: https://ikag.dev/blog/how-to-add-a-tool-to-chatgpt-claude-gemini - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-17 - Updated: 2026-10-17 - Reading time: 5 min - Publisher: IKAG ## Overview To add a tool to ChatGPT, Claude or Gemini, paste the tool’s MCP server address into the app’s settings and sign in to the service. In ChatGPT that is Plugins, then Add custom MCP server; in Claude, Customize, Connectors, then Add custom connector; in Gemini, Settings, Connected Apps, then Add a custom app. Cursor, VS Code, Claude Code and Codex take the same address in a settings page, a file or one command. Plan limits apply, and menus change often. ## Key points - Adding a tool to an AI app means pasting the tool’s MCP server address into the app and signing in to the service; we checked the steps below against each maker’s help pages on 17 October 2026. - ChatGPT adds custom MCP servers on the web only, with full support on Business, Enterprise and Edu and read and fetch permissions on Pro. - Claude lets every plan add custom connectors, one on the Free plan, while on Team and Enterprise an owner adds them for the organisation first. - Gemini adds custom apps on the web for personal Google Accounts in the US, for people aged 18 and over, in English, with Keep Activity on. - Cursor, VS Code, Claude Code and Codex take the same address through a settings page, a config file or one terminal command. To add a tool to ChatGPT, Claude or Gemini, you paste the tool’s MCP server address into the app’s settings and sign in to the service. Coding tools such as Cursor, VS Code, Claude Code and Codex take the same address through a settings page, a config file or one terminal command. Which plans can do it differs a lot. We checked every step below against each maker’s help pages on 17 October 2026. Menus change often, so if a button has moved, search the app’s help for “custom connector” or “MCP”. We use `https://example.com/mcp` as a stand-in; use the address your tool’s maker gives you. ## What do you need before you start? - **The server address.** The tool’s maker publishes it, usually in its docs. It has to be reachable over the internet: ChatGPT and Claude connect to servers remotely, so one that runs only on your own computer needs extra setup. - **An account with the service**, for the sign-in step. - **A plan that allows it.** See the table below. On work accounts, an admin may have to add the tool first. New to all this? Start with [what MCP is, in plain English](https://ikag.dev/blog/what-is-mcp-plain-english-guide). ## Which apps and plans let you add a tool? | App | Who can add one (October 2026) | Where | | ----------- | --------------------------------------------------------------------------------------- | --------------------------------------------- | | ChatGPT | Business, Enterprise and Edu in full, with admin control; Pro with read and fetch only | Web only | | Claude | Free (one connector), Pro and Max; on Team and Enterprise, an owner adds it first | Claude on the web, Cowork and Claude Desktop | | Gemini | Personal Google Accounts in the US, aged 18 and over, in English, with Keep Activity on | Added on the web, used on web and mobile | | Cursor | No plan limits listed; Enterprise admins can restrict which servers run | Cursor | | VS Code | No plan limits listed; organisations can manage access centrally | VS Code | | Claude Code | No plan limits listed on its MCP page | Terminal | | Codex | No plan limits listed on its MCP page | Terminal, shared with the ChatGPT desktop app | ## How do I add an MCP server to ChatGPT? On the web: 1. Open **Plugins**. 2. Choose the **+** button, then **Add custom MCP server**. 3. Enter a name and, if you like, a description. 4. Under **Connection**, paste the server address, such as `https://example.com/mcp`. 5. Set up authentication, so that you can sign in to the service. 6. Read the risk warning and tick **I understand and want to continue**. 7. Choose **Create as a plugin**, then review the tools it found. To use it, install the plugin it creates, then type @ in a new chat and pick it. According to [OpenAI’s help article](https://help.openai.com/en/articles/12584461-connecting-custom-mcp-servers-in-chatgpt), this works on the web only. On Business, only admins and owners can add a server; on Enterprise and Edu, admins can let selected members do it. Pro users can connect servers with read and fetch permissions only. ## How do I add a custom connector to Claude? On Pro or Max, and on Free, which allows one: 1. Go to **Customize**, then **Connectors**. 2. Choose **+ Add**, then **Add custom connector**. 3. Enter a name and the server address, then choose **Continue**. 4. Check the sign-in settings Claude detected and choose **Continue**. 5. Under **Authentication**, choose **Sign in now**, **Sign in when needed** or **No sign in**. 6. Leave the advanced options alone unless the tool’s docs say otherwise, choose **Add** and sign in to the service. Turn the connector on for a chat from the **+** button, then **Connectors**. On Team and Enterprise, an owner first adds it under **Organization settings**, then **Connectors**, and members then choose **Connect** ([Claude’s help centre](https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp)). ## How do I add a custom app to Gemini? 1. On a computer, go to gemini.google.com. 2. Open **Settings**, then **Connected Apps**. 3. Under **Custom apps**, choose **Add a custom app**. 4. Paste the server address and choose **Next**. 5. Follow the on-screen steps to sign in. Then type @ in a chat and pick the app, on the web or in the Gemini mobile app. [Google’s help page](https://support.google.com/gemini/answer/17209137) sets the conditions: you must be 18 or over and in the US, signed in with a personal Google Account rather than a work or school one, with Keep Activity on, using Gemini in English. Custom apps can only be added in the web app. ## How do I add an MCP server to Cursor or VS Code? **Cursor.** Install and manage servers from the **Customize** page, or add the server to `~/.cursor/mcp.json` to use it in every project: ```json { "mcpServers": { "example": { "url": "https://example.com/mcp" } } } ``` [Cursor’s docs](https://cursor.com/docs/mcp) say it supports sign-in (OAuth) for remote servers and, by default, asks for approval before it uses an MCP tool. **VS Code.** Run **MCP: Add Server** from the Command Palette, choose **HTTP** and paste the address, then give the server a name and choose where to save it. VS Code may ask you to trust the server when it first starts ([VS Code docs](https://code.visualstudio.com/docs/agent-customization/mcp-servers)). ## How do I add an MCP server to Claude Code or Codex? Both take one command in your terminal. In Claude Code: ```bash claude mcp add --transport http example https://example.com/mcp ``` Then type `/mcp` inside Claude Code and sign in. The server is saved for the current project; add `--scope user` to use it in all your projects ([Claude Code docs](https://code.claude.com/docs/en/mcp)). In Codex: ```bash codex mcp add example --url https://example.com/mcp codex mcp login example ``` The second command signs you in, for servers that use OAuth. Codex keeps the server in `~/.codex/config.toml`, which the ChatGPT desktop app, the Codex CLI and the IDE extension share ([OpenAI’s MCP docs](https://learn.chatgpt.com/docs/extend/mcp?surface=cli)). ## Can agents like ChatGPT Dots and Grok Bot use these tools? Partly. OpenAI says ChatGPT Dots can use [the plugins you’ve set up in ChatGPT](https://help.openai.com/en/articles/20001529-dots-privacy-security-and-safety-faqs), so a server added there can reach your agent, though [the apps available to it may differ](https://help.openai.com/en/articles/20001530-getting-started-with-your-dot). Grok Bot adds a server when you ask it in chat, [according to Cursor’s support team](https://forum.cursor.com/t/grokbot-custom-connectors/169965). [Meta’s help page](https://www.meta.com/help/artificial-intelligence/1687253048996149/) describes custom connectors that Muse builds from a service’s API, and doesn’t mention MCP. We compare all three in [ChatGPT Dots, Grok Bot and Meta Muse explained](https://ikag.dev/blog/chatgpt-dots-grok-bot-meta-muse-explained). ## What if it doesn’t work? - **Check the address.** It must be the MCP server address, not the service’s homepage, and reachable over the internet. - **Check your plan and role.** On work plans, an admin may have to add or approve the tool. - **Sign in again.** Sign-ins can expire; reconnect from the app’s connector settings. - **Ask an admin to refresh it.** In ChatGPT Business and Enterprise, a server’s tools are fixed when an admin approves it, so later changes need an admin to refresh them. Only add servers you trust. For what to check first, see [what to ask before you connect an agent](https://ikag.dev/blog/questions-before-you-connect-an-agent). ## Where IKAG fits IKAG is designed to connect exactly this way: one MCP link in apps like these, then a sign-in the first time your AI needs it. [How pay per call is meant to work](https://ikag.dev/#pricing) is on the homepage; no prices have been set. IKAG is opening to early users and data partners first. ## Questions ### Can I add an MCP server to ChatGPT Plus? OpenAI’s help article lists full custom MCP support for Business, Enterprise and Edu, and read and fetch permissions for Pro. It doesn’t mention Plus or Free, so check the Plugins page in your own account. ### How many custom connectors can I add to Claude for free? One. Claude’s help centre says Free users are limited to one custom connector. Pro and Max can add more, and on Team and Enterprise an owner adds the connector first. ### Why can’t I add a custom app in Gemini? Google says you must be 18 or over and in the US, signed in with a personal Google Account, with Keep Activity on, using Gemini in English. Custom apps can only be added in the Gemini web app. ### Can I add an MCP server from my phone? Not in ChatGPT or Gemini. ChatGPT adds custom MCP servers on the web only, and Gemini adds custom apps only in its web app, though you can then use them in the Gemini mobile app. ## Sources - [OpenAI Help Center: Connecting custom MCP servers in ChatGPT (updated October 2026)](https://help.openai.com/en/articles/12584461-connecting-custom-mcp-servers-in-chatgpt) - [Claude Help Center: Get started with custom connectors using remote MCP (checked 17 October 2026)](https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp) - [Google Gemini Apps Help: Connect & manage custom apps for Gemini Apps (checked 17 October 2026)](https://support.google.com/gemini/answer/17209137) - [Cursor Docs: Model Context Protocol (MCP) (checked 17 October 2026)](https://cursor.com/docs/mcp) - [Visual Studio Code: Add and manage MCP servers in VS Code (14 October 2026)](https://code.visualstudio.com/docs/agent-customization/mcp-servers) - [Claude Code Docs: Connect Claude Code to tools via MCP (checked 17 October 2026)](https://code.claude.com/docs/en/mcp) - [OpenAI: Model Context Protocol in Codex, ChatGPT Learn (checked 17 October 2026)](https://learn.chatgpt.com/docs/extend/mcp?surface=cli) - [OpenAI Help Center: Dots privacy, security, and safety FAQs (updated October 2026)](https://help.openai.com/en/articles/20001529-dots-privacy-security-and-safety-faqs) - [OpenAI Help Center: Getting started with your dot (updated October 2026)](https://help.openai.com/en/articles/20001530-getting-started-with-your-dot) - [Cursor Community Forum: Grokbot custom connectors, answer from Cursor support (30 August 2026)](https://forum.cursor.com/t/grokbot-custom-connectors/169965) - [Meta Help Centre: How Muse works with Connectors (checked 17 October 2026)](https://www.meta.com/help/artificial-intelligence/1687253048996149/) --- # Local business data for AI agents: sources and rules > How an AI agent can list local businesses with ratings, reviews, websites and opening hours: map platforms, open data and business websites, their terms and freshness. - URL: https://ikag.dev/blog/local-business-data-for-ai-agents - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-17 - Updated: 2026-10-17 - Reading time: 5 min - Publisher: IKAG ## Overview An AI agent lists local businesses by calling a places source: a map platform’s API for ratings, review counts, websites and opening hours, open map data such as OpenStreetMap or Overture for records you may keep, and each business’s own website to confirm details that change. Every source has terms. Google’s forbid copying business names, addresses and reviews out of its services, and OpenStreetMap’s require credit and share-alike. A good answer gives the source and date of every value and flags what it couldn’t confirm. ## Key points - No single source has everything: map platforms hold ratings and reviews, open map data holds records you can reuse, and a business’s own website is often the best check on its opening hours. - Google Maps Platform’s terms forbid copying and saving business names, addresses or user reviews outside its services, but place IDs may be stored indefinitely. - OpenStreetMap data is free to reuse under the Open Database License if you credit its contributors and share adapted data under the same licence. It does not collect reviews or customer ratings. - A rating means little without its review count, source and date. Google says it doesn’t verify reviews but removes fake content it identifies, and UK law has banned fake reviews since 6 April 2025. - A good answer gives each value’s source and date, marks businesses that are closed or unconfirmed, and lists what it couldn’t find. Ask your AI for every independent gym in a city, with ratings, websites and opening hours, and it needs a places source. A model learned from data gathered before its training ended, and it was never a directory: businesses open, close and change their hours all the time. The agent has to look each business up, and a good answer says where each value came from. ## Where does local business data come from? Four kinds of source, each good at different things. **Map platforms.** Google’s Places API returns a place’s name, address, [rating, number of ratings, website, opening hours and business status](https://developers.google.com/maps/documentation/places/web-service/data-fields), plus [up to five reviews](https://developers.google.com/maps/documentation/places/web-service/reference/rest/v1/places). It needs a billing account and a key, and each request is [billed at the highest price tier of the fields you ask for](https://developers.google.com/maps/documentation/places/web-service/usage-and-billing): ratings, websites and hours sit in a higher tier than names and addresses, and reviews in a higher one still. **Open map data.** [OpenStreetMap](https://www.openstreetmap.org/copyright) is built by a community of contributors and licensed under the Open Database License. It has standard tags for a place’s [opening hours](https://wiki.openstreetmap.org/wiki/Key:opening_hours) and [website](https://wiki.openstreetmap.org/wiki/Key:website), but by [its own rules](https://wiki.openstreetmap.org/wiki/Verifiability) it holds no reviews or customer ratings, because another mapper couldn’t check an opinion. The Overture Maps Foundation’s [places data](https://docs.overturemaps.org/guides/places/) holds tens of millions of places from sources including Meta, Microsoft and Foursquare, under permissive licences, released monthly. Overture says plainly that it contains duplicates, a high junk rate and many missing fields, and gives each place a confidence score to filter on. **Official registers.** Some ratings are inspections, not opinions. In the UK, the Food Standards Agency publishes [food hygiene ratings](https://ratings.food.gov.uk/open-data) for restaurants, pubs, cafés, takeaways and food shops through a free API, each reflecting the standards found on the date of inspection. **The businesses’ own websites.** A business’s site is often the most current word on its hours and services, and a site can describe them in a standard format: [schema.org’s LocalBusiness](https://schema.org/LocalBusiness) has fields for opening hours. Reading sites means following each site’s terms and its [robots.txt](https://www.rfc-editor.org/rfc/rfc9309.html). | Source | Good for | Terms | Freshness | | ------------------------- | --------------------------------------- | -------------------------------------- | --------------------------------------- | | Map platform API | Ratings, review counts, websites, hours | Credit required; no copying out | Shows if a place has closed | | OpenStreetMap | Places, hours, websites | ODbL: credit and share-alike | Can record when hours were last checked | | Overture places | A large list of places you can reuse | Permissive open licences | Monthly releases; a confidence score | | Food hygiene ratings (UK) | Official hygiene ratings | The Food Standards Agency’s conditions | The date of inspection | | The business’s website | The latest hours and services | Site terms and robots.txt | As current as the business keeps it | ## What do the terms allow? This is where a request that ends “in a sheet” gets interesting. [Google Maps Platform’s terms](https://cloud.google.com/maps-platform/terms) say customers will not scrape or export its content for use outside its services, and give copying and saving “business names, addresses, or user reviews” as an example. They also rule out caching, except where Google allows it, and using Maps content to train AI models. Google’s [general terms](https://policies.google.com/terms) separately forbid automated access that ignores the robots.txt instructions on its pages. What you may keep is narrow: Google’s [Places API policies](https://developers.google.com/maps/documentation/places/web-service/policies) let you store place IDs indefinitely, and its [service terms](https://cloud.google.com/maps-platform/terms/maps-service-terms) allow coordinates for up to 30 days. Content you show must credit Google Maps, and each review must credit its author and link to the original. Customers with a billing address in the European Economic Area have separate Maps terms. OpenStreetMap makes the opposite trade. You may copy and adapt its data, but you must credit OpenStreetMap and its contributors, and anything you build on it and distribute must use the same licence. It also says it can’t provide a free map API for third parties, so large jobs work from a downloaded copy or another provider. None of this is legal advice. The practical rule is to know which source every value came from, because each value carries its source’s conditions with it. ## How fresh is the data? Businesses close and hours change, so every value needs a date. - **Status.** Google marks each place as operational, temporarily closed or permanently closed, and operational doesn’t mean open right now. - **Opening hours.** OpenStreetMap mappers can record when they last checked a place’s hours with a [check_date tag](https://wiki.openstreetmap.org/wiki/Key:check_date). A business’s own website may be newer than either. - **Ratings and reviews.** Google [says](https://developers.google.com/maps/documentation/places/web-service/policies) it doesn’t verify reviews but checks for and removes fake content when it finds it. In the UK, [fake reviews have been banned since 6 April 2025](https://www.gov.uk/government/news/fake-reviews-and-sneaky-hidden-fees-banned-once-and-for-all), and websites must take steps to prevent and remove them. The same average from a dozen reviews and from several hundred is not the same finding, so keep the count. ## What should a good answer include? Here is the kind of request you might give your AI. Illustrative: > “Every independent gym in Manchester rated under 4.2, with their website and review count, in a sheet.” For each business, a good answer gives: - **Identity:** name, address and the source’s own ID, such as a Google place ID, so the row can be refreshed later. - **The rating in context:** the rating, the number of reviews behind it, the source and the date. - **Website and hours:** with where each came from and when it was last confirmed. - **Status:** open, temporarily closed or permanently closed. - **How “independent” was judged:** no source holds that field, so the agent should say what it used, such as finding no other branches. - **Credit:** the attribution each source requires. - **Gaps, stated:** “no website listed” or “hours not confirmed” as facts, not blank cells. It should also respect the source’s terms: an answer you read once is one thing, a spreadsheet you keep is another. For checking the result, see [how to check where an AI answer came from](https://ikag.dev/blog/how-to-check-ai-answers), and for the ways an agent reaches outside data, [how AI agents get live data](https://ikag.dev/blog/how-ai-agents-get-live-data). ## Where IKAG fits Local business data is one of the next kinds of tools IKAG is building. Your AI, connected once, would ask for the list, and IKAG is designed to call a places source within its terms and return where each value came from, where the source supports it. You would pay per call instead of a subscription to a listings tool, and no prices have been set. IKAG is opening to early users and data partners first. ## Questions ### Can an AI agent read Google Maps reviews? Through Google’s Places API it can get a place’s rating, its number of ratings and up to five reviews, billed per request. Google’s terms require you to credit each review’s author and forbid copying and saving reviews outside its services. ### Is scraping Google Maps allowed? Not on Google’s terms. Its Maps Platform terms forbid scraping, exporting or caching Maps content outside its services, including copying and saving business names, addresses and user reviews, and its general terms forbid automated access that ignores its robots.txt instructions. Use the official API within its terms, or open data; this is not legal advice. ### Can I use OpenStreetMap data commercially? Yes, under the Open Database License: credit OpenStreetMap and its contributors, and share any adapted database you distribute under the same licence. OpenStreetMap says it cannot offer a free map API for third parties, so large jobs work from a copy of the data or another provider. ### How fresh is local business data? It depends on the source and the field. Opening hours change most often: OpenStreetMap can record when a mapper last checked them, Google marks places as temporarily or permanently closed, and a business’s own website is often the latest word. ## Sources - [Google for Developers: Place data fields, Places API (New) (last updated 15 October 2026)](https://developers.google.com/maps/documentation/places/web-service/data-fields) - [Google for Developers: REST resource places, Places API (New) (last updated 6 August 2026)](https://developers.google.com/maps/documentation/places/web-service/reference/rest/v1/places) - [Google for Developers: Places API usage and billing (last updated 15 October 2026)](https://developers.google.com/maps/documentation/places/web-service/usage-and-billing) - [Google for Developers: Policies and attributions for Places API (last updated 15 October 2026)](https://developers.google.com/maps/documentation/places/web-service/policies) - [Google: Google Maps Platform Terms of Service (last modified 26 August 2026)](https://cloud.google.com/maps-platform/terms) - [Google: Google Maps Platform Service Specific Terms (last modified 15 October 2026)](https://cloud.google.com/maps-platform/terms/maps-service-terms) - [Google: Terms of Service, UK version (effective 30 July 2026)](https://policies.google.com/terms) - [OpenStreetMap: Copyright and licence (checked 17 October 2026)](https://www.openstreetmap.org/copyright) - [OpenStreetMap Wiki: Verifiability (last edited 28 November 2024)](https://wiki.openstreetmap.org/wiki/Verifiability) - [OpenStreetMap Wiki: Key opening_hours (last edited 5 September 2026)](https://wiki.openstreetmap.org/wiki/Key:opening_hours) - [OpenStreetMap Wiki: Key website (checked 17 October 2026)](https://wiki.openstreetmap.org/wiki/Key:website) - [OpenStreetMap Wiki: Key check_date (last edited 18 May 2026)](https://wiki.openstreetmap.org/wiki/Key:check_date) - [Overture Maps Foundation: Places guide (September 2026 release)](https://docs.overturemaps.org/guides/places/) - [Food Standards Agency: UK food hygiene rating data API (checked 17 October 2026)](https://ratings.food.gov.uk/open-data) - [Schema.org: LocalBusiness (version 30.1, 16 September 2026)](https://schema.org/LocalBusiness) - [RFC 9309: Robots Exclusion Protocol (September 2022)](https://www.rfc-editor.org/rfc/rfc9309.html) - [GOV.UK: Fake reviews and sneaky hidden fees banned once and for all (6 April 2025)](https://www.gov.uk/government/news/fake-reviews-and-sneaky-hidden-fees-banned-once-and-for-all) --- # Pay per call vs monthly plans for AI tools > Monthly plans suit steady use and pay per call suits bursts. How they compare for AI tools: quotes before a call, spending limits, receipts and your own keys. - URL: https://ikag.dev/blog/pay-per-call-vs-monthly-plans-for-ai-tools - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-17 - Updated: 2026-10-17 - Reading time: 6 min - Publisher: IKAG ## Overview A monthly plan charges for access whether you use it or not; pay per call charges for each request a tool handles. Plans suit steady, heavy use. Pay per call suits occasional or bursty use, which is often how AI agents use tools, provided you can see the price before a call runs, cap what it can spend and get a receipt for each call. ## Key points - A monthly plan charges for access; pay per call charges for each request. Which is better depends on how often, and how evenly, you use a tool. - Usage pricing has its own risk: in a survey of 218 IT leaders for Zylo’s 2026 index, 78% reported unexpected charges tied to consumption-based or AI pricing in the past year. - Pay per call is safe to use when three things come with it: a quote before each call, a spending limit that stops calls, and a receipt for each call. - With your own provider key, that provider’s plan and charges still apply, and a service that routes your calls may charge separately for its part. - IKAG is opening to early users and data partners first. It is designed around pay per call, and no prices have been set. There are two basic ways to pay for the tools your AI uses. A monthly plan charges for access, whether you use it or not. Pay per call charges for each request the tool handles. Neither is right for everything, and the difference matters more now that AI agents, not just people, make the calls. The short answer: pick a plan for a tool you use every day, and pay per call for tools you use occasionally or in bursts. Pay per call needs three protections to be safe: a quote before the call, a limit it can’t pass and a receipt afterwards. ## What is the difference between pay per call and a monthly plan? A **monthly plan** sells access for a period. You pay the same on the first of the month whether you made one request or a thousand, often up to an allowance, with a higher tier when you need more. **Pay per call** sells each request: one lookup, one page read, one video made. A quiet month costs little and a busy one costs more. It is one form of usage-based pricing, alongside pricing per token, per minute or per gigabyte. Some services mix the two: a plan that includes a set amount of use, with a charge for anything above it. ## When does a monthly plan make sense? When your use is steady and heavy. If you open a tool every day, a flat fee is easy to budget and can work out cheaper than paying for each request. Plans also suit teams that want everyone on the same tool at a known cost per seat. The catch is that a plan charges for the whole month. [Zylo’s 2026 index](https://zylo.com/news/2026-saas-management-index) found that organisations leave 36% of their software licences unused on average, and [subscription fatigue](https://ikag.dev/blog/subscription-fatigue-tools-you-use-twice) covers what that looks like for companies and households. ## When does pay per call make sense? When your use is occasional, bursty or uncertain, which is often how AI agents use tools. An agent might need lead data on Monday, AI video on Wednesday and web data on Friday, a few calls each, then nothing for a fortnight. A month of access to each tool for that pattern is mostly paying for days it sat idle. Pay per call also lowers the cost of trying a tool. Your AI can test it on one real task before you commit to anything. ## How do they compare? | Question | Monthly plan | Pay per call | Your own key through a service | | ------------------------ | ----------------------- | --------------------------- | ---------------------------------------- | | What you pay for | Access for the month | Each call that runs | Your provider plan, plus any service fee | | Best for | Daily, steady use | Occasional or bursty use | Tools you already pay for | | A month you don’t use it | The full fee | Little or nothing | Your provider plan’s fee | | Main risk | Paying for idle access | A bill that grows with use | Two sets of charges to track | | How you stay in control | Cancel or downgrade | Quotes, limits and receipts | Your provider’s limits and the service’s | | Leaving | Cancel before it renews | Stop making calls | Remove the key | ## What can go wrong with usage pricing? The bill moves with use, and use can surprise you. In a survey of 218 IT leaders for [Zylo’s 2026 SaaS Management Index](https://zylo.com/news/2026-saas-management-index), 78% reported unexpected charges tied to consumption-based or AI pricing in the past 12 months, and 61% had to cut projects because software costs rose unexpectedly. **Chart: IT leaders who met surprise software costs in the past year** | Item | Value | | --- | --- | | Unexpected usage or AI charges | 78% | | Had to cut projects | 61% | Survey of 218 IT leaders for Zylo’s 2026 SaaS Management Index: unexpected charges tied to consumption-based or AI pricing, and projects cut because software costs rose unexpectedly. Source: [Zylo: 2026 SaaS Management Index finds AI-native app adoption is surging (29 January 2026)](https://zylo.com/news/2026-saas-management-index). Agents add a risk of their own: one stuck in a loop can repeat the same paid call many times before anyone notices. So pay per call only beats a plan if you can see, cap and check what it spends. That takes three things. ## Why see the price before a call runs? Because a price you learn at the end of the month is a surprise, not a price. A quote before the call lets you, or your AI, decide whether a request is worth it. The idea is already being tried on the open web. Cloudflare’s [pay per crawl](https://blog.cloudflare.com/introducing-pay-per-crawl/), introduced in private beta in July 2025, revives a mostly forgotten part of the web, the 402 “payment required” response, to tell AI crawlers the price of a page before they take it. A crawler can then agree to that price, or say up front the most it will pay, and each paid response confirms what was charged. The pattern carries over to any per-call tool: price first, then the call, then a record of the charge. ## What makes a spending limit real? A real limit stops calls when it is reached. An alert only tells you. The difference is easy to miss. Anthropic lets organisations [set their own spend limit](https://platform.claude.com/docs/en/api/rate-limits) on its API, below their tier’s monthly cap; once usage reaches it, requests are refused until access resumes or the limit is raised. Google Cloud’s [budget documentation](https://docs.cloud.google.com/billing/docs/how-to/budgets) warns that an alerts-only budget doesn’t automatically cap usage or spending, and points to a separate spend cap where a service supports one. Before you let an agent use a paid tool on its own, check which kind of limit you have, and set one per tool or per task where you can. ## What should a receipt for each call show? Enough to check the charge against the work: - **The tool and the request:** what ran, and for which task. - **The price:** what was quoted and what was charged. - **The time:** when the call ran. - **The outcome:** whether it worked. A failed call should be marked as failed, never presented as an answer. Itemised receipts also show which tools earn their keep. If one tool takes most of your spending, a plan for that tool alone may be cheaper. [What to ask before you connect an agent](https://ikag.dev/blog/questions-before-you-connect-an-agent) has more questions worth asking about cost. ## What happens when you bring your own key? Bringing your own key means connecting your own account with a provider, so calls run under your plan with them. You keep your limits, your data terms and any plan you’ve already paid for. The provider still bills you as before, and a service in between may charge separately for its part. Why each provider insists on an account and a key at all is its own story: [why services sell access through accounts and plans](https://ikag.dev/blog/do-you-need-a-subscription-for-every-ai-tool). OpenRouter, a service that routes requests to many AI models, is a published example. Using your own provider keys, its documentation says, gives you [direct control over rate limits and costs](https://openrouter.ai/docs/guides/overview/auth/byok) through your provider account, and OpenRouter charges a fee of 5% of what the same model would normally cost through it, beyond an allowance that depends on your plan. Expect two sets of charges, and check both. ## Where IKAG fits IKAG is designed so the AI you already use pays per call for the tools it needs, with no subscription to each service. For the tools IKAG manages, you would see a quote before a call runs, set a limit it can’t pass and get each call itemised on a receipt; with your own provider account, that provider’s plan and charges still apply, and IKAG may charge separately for its part. No prices have been set, and they will be published before IKAG charges anyone. IKAG is opening to early users and data partners first; see [how pricing is designed](https://ikag.dev/#pricing). ## Questions ### Is pay per call cheaper than a monthly subscription? It depends on how much you use the tool. For occasional or bursty use, paying per call can cost less than a month of access you barely use; for steady, heavy use, a monthly plan is often better value. ### How do I stop an AI agent running up a bill? Use a spending limit that stops calls when it is reached, not just an alert, and prefer tools that quote a price before each call. Then check the receipts for calls you did not expect, such as an agent repeating the same call. ### What happens to pricing if I use my own API key? Calls run on your account with that provider, so its plan, limits and charges still apply. A service that routes those calls may charge separately for its part. ### How will IKAG pricing work? Pay per call is the direction: for the tools IKAG manages, you would see a quote before a call runs and pay for the calls your AI makes, with no subscription to each service. No prices have been set, and with your own provider account, that provider’s plan and charges still apply. IKAG is opening to early users and data partners first. ## Sources - [Zylo: 2026 SaaS Management Index finds AI-native app adoption is surging (29 January 2026)](https://zylo.com/news/2026-saas-management-index) - [Cloudflare: Introducing pay per crawl: Enabling content owners to charge AI crawlers for access (1 July 2025)](https://blog.cloudflare.com/introducing-pay-per-crawl/) - [Anthropic, Claude Platform docs: Rate limits and spend limits (accessed 17 October 2026)](https://platform.claude.com/docs/en/api/rate-limits) - [Google Cloud documentation: Create, edit, or delete budgets and budget alerts (updated 13 October 2026)](https://docs.cloud.google.com/billing/docs/how-to/budgets) - [OpenRouter docs: Bring your own provider API keys (accessed 17 October 2026)](https://openrouter.ai/docs/guides/overview/auth/byok) --- # What is MCP? A plain-English guide for AI users > MCP, the Model Context Protocol, lets ChatGPT, Claude, Gemini and other AI apps use outside tools. What it is, how connecting works and how to stay safe. - URL: https://ikag.dev/blog/what-is-mcp-plain-english-guide - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-17 - Updated: 2026-10-17 - Reading time: 5 min - Publisher: IKAG ## Overview MCP, the Model Context Protocol, is an open standard for connecting AI apps to outside tools and data. A service publishes an MCP server at a web address; you add that address to Claude, ChatGPT, Gemini, Cursor or another app, sign in to the service, and your AI can use its tools. MCP standardises the connection. It doesn’t choose the right source for a question, and it doesn’t make a server safe. ## Key points - MCP, the Model Context Protocol, is an open standard for connecting AI apps to outside tools and data, introduced by Anthropic in November 2024 and given to the Agentic AI Foundation in December 2025. - Connecting an MCP server on the web usually takes a link and a sign-in: you paste the server’s address into your AI app, then approve access on the service’s own sign-in page. - Claude, ChatGPT, Gemini, Grok, Cursor, VS Code, Claude Code and Codex can all connect to MCP servers, within limits set by each app and plan. - MCP standardises the connection; it doesn’t pick the right source for a question or make a server trustworthy, so connect only servers you trust and grant the least access a task needs. MCP, short for Model Context Protocol, is a shared way for AI apps to use outside tools and data. A service publishes an MCP server at a web address. You add that address to Claude, ChatGPT, Gemini, Cursor or another app, sign in, and your AI can use the service’s tools in a conversation. The [official introduction](https://modelcontextprotocol.io/docs/getting-started/intro) compares it to “a USB-C port for AI applications”: one standard plug instead of a different cable for every device. You don’t need to build anything to use it. You need the link, an account with the service, and an app and plan that allow custom connections. ## Why does MCP matter? Before MCP, connecting an AI app to a service usually meant a custom integration for that one pair. When Anthropic [introduced MCP](https://www.anthropic.com/news/model-context-protocol) on 25 November 2024, it described the problem as each new data source needing its own custom implementation. With a shared standard, a service builds one server and many apps can use it. It is now widely supported. Claude, ChatGPT, Gemini, Grok, Cursor, VS Code, Claude Code and Codex all document how to add an MCP server. On 9 December 2025 Anthropic [donated MCP to the Agentic AI Foundation](https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation), a fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI, and counted more than 10,000 active public MCP servers at the time. For you, that means choice. The tools your AI can use are no longer only the ones its maker built in. ## How does connecting a server work? For a server on the web, it takes a link and then a sign-in. 1. **Get the address.** The service publishes its MCP server address, for example `https://example.com/mcp`. 2. **Add it to your app.** Each app has a settings page for this, under its own name (see the table below). 3. **Sign in.** Your browser opens the service’s own sign-in page. You approve what the app may do, and the app receives a token instead of your password. 4. **Ask.** Your AI now sees the server’s tools and decides when to use them. In ChatGPT and Gemini, typing @ and picking the tool makes sure it is used. The sign-in step is usually OAuth. MCP’s [authorisation rules](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization) build on OAuth 2.1 for servers on the web and make sign-in optional, so some servers need none and others take an API key instead. [Claude’s help centre](https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp) sums it up: you grant specific permissions without Claude ever seeing your password, and you can revoke them later in Claude’s settings or in the service’s own. Coding tools can also run MCP servers on your own computer, started by a command. Those can run code on your machine, so [VS Code’s docs](https://code.visualstudio.com/docs/agent-customization/mcp-servers) advise adding them only from trusted sources. ## Where does MCP work today? The main AI apps support it, but each calls it something different. | App | What it calls an MCP server | Where you add one | | ------------------ | --------------------------------------- | ----------------------------------------- | | Claude | Custom connector | Customize, then Connectors | | ChatGPT | Custom MCP server, added as a plugin | Plugins, on the web | | Gemini | Custom app | Settings, then Connected Apps, on the web | | Grok | Custom connector (“Bring Your Own MCP”) | grok.com/connectors, then New Connector | | Cursor | MCP server | The Customize page, or a settings file | | VS Code | MCP server | The MCP: Add Server command | | Claude Code, Codex | MCP server | One terminal command | Plans decide who can add one. [ChatGPT](https://help.openai.com/en/articles/12584461-connecting-custom-mcp-servers-in-chatgpt) gives full support on Business, Enterprise and Edu and read and fetch permissions only on Pro; Claude’s Free plan allows one custom connector; [Gemini’s custom apps](https://support.google.com/gemini/answer/17209137) are for personal accounts in the US. On [Grok](https://docs.x.ai/grok/connectors), connectors are available to all users, with an admin adding them first on Business and Enterprise. The full steps and limits are in [how to add a tool to ChatGPT, Claude and Gemini](https://ikag.dev/blog/how-to-add-a-tool-to-chatgpt-claude-gemini). The newer agents connect too, each in its own way: see [ChatGPT Dots, Grok Bot and Meta Muse explained](https://ikag.dev/blog/chatgpt-dots-grok-bot-meta-muse-explained). ## What doesn’t MCP do? MCP standardises the plug, not what’s behind it. - **It doesn’t choose the right source.** Your AI picks tools from their names and descriptions. Connect ten servers and it chooses between ten sets of tools; nothing in the protocol says which source is right for a question. We compare two answers to that in [tool routers and data networks](https://ikag.dev/blog/tool-routers-and-data-networks-for-ai-agents). - **It doesn’t make answers correct.** A tool returns what its service returns. Whether that is current and accurate depends on the service. - **It doesn’t handle accounts or payment.** Each server brings its own account, pricing and terms. A paid service with an MCP server still needs its plan. - **It can’t enforce safety.** The [specification](https://modelcontextprotocol.io/specification/2026-07-28) says MCP cannot enforce its security principles at the protocol level. Consent, permissions and approvals are up to the apps and the servers. ## How do you stay safe with MCP servers? Claude, ChatGPT, Gemini, Cursor and VS Code give the same first rule in their help pages: only connect servers you trust. A few habits help. - **Know who runs the server.** Prefer one published by the service itself, on its own domain. - **Read the sign-in screen.** It shows what the server wants to access. Claude’s help centre advises limiting those permissions where you can and refusing ones that seem unnecessary. - **Tell reading from acting.** Some tools only fetch information; others create, change, send or delete. Gemini asks you to confirm write actions, and ChatGPT limits Pro users to read and fetch. - **Keep approvals on.** Claude suggests choosing “Allow always” only for servers and tools you trust to run unsupervised, and [Cursor](https://cursor.com/docs/mcp) asks before using an MCP tool by default. - **Watch for prompt injection.** A malicious server can include hidden instructions meant to make your AI do something you didn’t ask. Anthropic and OpenAI both warn about it. - **Tidy up.** Disconnect servers you no longer use. [OpenAI](https://help.openai.com/en/articles/20001529-dots-privacy-security-and-safety-faqs) and [Meta](https://www.meta.com/help/artificial-intelligence/1687253048996149/) note that disconnecting stops new access but doesn’t delete what their agents already took in. For a longer checklist, see [what to ask before you connect an agent](https://ikag.dev/blog/questions-before-you-connect-an-agent). ## Where IKAG fits IKAG is building one MCP server that gives the AI people already use the paid tools and data they can’t reach on their own, such as lead data, AI video, web data and company records, paid per call instead of a subscription to each service, though no prices have been set. Connecting is designed to work as this guide describes: one link, then a sign-in, and [how to add a tool](https://ikag.dev/blog/how-to-add-a-tool-to-chatgpt-claude-gemini) has the steps for each app. IKAG is opening to early users and data partners first. ## Questions ### What does MCP stand for? Model Context Protocol. It is an open-source standard for connecting AI applications to external systems such as data sources, tools and workflows. ### Who owns MCP? Anthropic introduced MCP on 25 November 2024. On 9 December 2025 it donated the protocol to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI. ### Do I need to code to use MCP? No. In chat apps such as Claude, ChatGPT and Gemini you paste a link into a settings page and sign in. Coding tools such as Claude Code and Codex take a one-line terminal command instead. ### Is it safe to connect an MCP server? It depends on the server. MCP’s own specification says the protocol cannot enforce its security principles, so connect only servers you trust, read what each one asks to access, and keep approvals on for actions that change or send things. ## Sources - [Model Context Protocol: What is the Model Context Protocol (MCP)? (checked 17 October 2026)](https://modelcontextprotocol.io/docs/getting-started/intro) - [Anthropic: Introducing the Model Context Protocol (25 November 2024)](https://www.anthropic.com/news/model-context-protocol) - [Anthropic: Donating the Model Context Protocol and establishing the Agentic AI Foundation (9 December 2025)](https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation) - [Model Context Protocol specification, version 2026-07-28: overview and security principles](https://modelcontextprotocol.io/specification/2026-07-28) - [Model Context Protocol specification, version 2026-07-28: Authorization](https://modelcontextprotocol.io/specification/2026-07-28/basic/authorization) - [Claude Help Center: Get started with custom connectors using remote MCP (checked 17 October 2026)](https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp) - [OpenAI Help Center: Connecting custom MCP servers in ChatGPT (updated October 2026)](https://help.openai.com/en/articles/12584461-connecting-custom-mcp-servers-in-chatgpt) - [Google Gemini Apps Help: Connect & manage custom apps for Gemini Apps (checked 17 October 2026)](https://support.google.com/gemini/answer/17209137) - [x.ai: Connectors in web, iOS, and Android (6 May 2026)](https://x.ai/news/grok-connectors) - [Grok docs: Connectors (updated 10 October 2026)](https://docs.x.ai/grok/connectors) - [Cursor Docs: Model Context Protocol (MCP) (checked 17 October 2026)](https://cursor.com/docs/mcp) - [Visual Studio Code: Add and manage MCP servers in VS Code (14 October 2026)](https://code.visualstudio.com/docs/agent-customization/mcp-servers) - [Claude Code Docs: Connect Claude Code to tools via MCP (checked 17 October 2026)](https://code.claude.com/docs/en/mcp) - [OpenAI: Model Context Protocol in Codex, ChatGPT Learn (checked 17 October 2026)](https://learn.chatgpt.com/docs/extend/mcp?surface=cli) - [OpenAI Help Center: Dots privacy, security, and safety FAQs (updated October 2026)](https://help.openai.com/en/articles/20001529-dots-privacy-security-and-safety-faqs) - [Meta Help Centre: How Muse works with Connectors (checked 17 October 2026)](https://www.meta.com/help/artificial-intelligence/1687253048996149/) --- # Do AI summaries send people to their sources? > Pew Research found Google users clicked a result in 8% of visits with an AI summary and 15% without. What that means for answers you need to check. - URL: https://ikag.dev/blog/do-ai-summaries-send-people-to-sources - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-15 - Updated: 2026-10-15 - Reading time: 3 min - Publisher: IKAG ## Overview Mostly not. In Pew Research Center’s study of 900 US adults’ browsing in March 2025, people clicked a traditional search result in 8% of visits to a Google results page that showed an AI summary, against 15% when it did not, and clicked a link inside the summary in 1% of visits. If people rarely click through, the evidence has to travel with the answer itself. ## Key points - In Pew’s March 2025 browsing data, Google users clicked a traditional result in 8% of visits that showed an AI summary and in 15% of visits that did not. - They clicked a link inside the AI summary in 1% of visits to pages with one, and ended their browsing session on 26% of those pages, against 16% without a summary. - About 18% of the Google searches in the study produced an AI summary. - When few people click through, the source record has to travel with the answer: the record, its date and what it does not say, not only a link. AI summaries now sit above the links on many search pages, and assistants answer questions without sending anyone anywhere. That is convenient. It also changes who checks the answer. The best data on what people actually do comes from a Pew Research Center study published in July 2025, and its numbers are worth knowing if you build or rely on AI answers. ## What did Pew find? Pew analysed the browsing of 900 US adults in March 2025, covering 68,879 unique Google searches. About 18% of those searches produced an AI summary at the top of the results. When a results page showed an AI summary, people clicked a traditional search result in 8% of visits. When it did not, they clicked one in 15% of visits, nearly twice as often. Links inside the summary itself were clicked in 1% of visits to pages that had one. **Chart: Share of visits to a Google results page with a click, March 2025** | Item | Value | | --- | --- | | Result click, no summary | 15% | | Result click, with summary | 8% | | Click inside the summary | 1% | 900 US adults, browsing data from March 2025. “Summary” means a Google AI Overview on the results page. Source: [Pew Research Center: Google users are less likely to click on links when an AI summary appears (22 July 2025)](https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/). People were also more likely to stop. Pew found that 26% of results pages with an AI summary were the last page of a browsing session, against 16% of pages without one. **Chart: Results pages after which people ended their browsing session** | Item | Value | | --- | --- | | Without an AI summary | 16% | | With an AI summary | 26% | Share of visits to Google results pages that were the last page of the browsing session. Source: [Pew Research Center: Google users are less likely to click on links when an AI summary appears (22 July 2025)](https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/). ## Why fewer clicks matter for checking answers A citation only helps if someone follows it. When summaries answer the question on the page, fewer people open the source, and the source’s job shifts from “where to read more” to “what this answer rests on”. That would be fine if citations were reliable. They are not always. In a March 2025 test by the Tow Center at Columbia Journalism Review, eight AI search tools asked to identify the source of quoted news excerpts got more than 60% of 1,600 queries wrong. That measured one task, not overall accuracy, but it shows that a link next to an answer is a claim in its own right. ## What should travel with the answer instead If people will not go to the record, the record has to come to them. For each value an answer depends on, that means: - **The source and the record**: not just a site, but the record that states the value, with its identifier. - **The record’s own date**: when the source last attested the fact, which is often not when it was fetched. - **What the record does not say**: the questions it leaves open, stated as unknowns rather than filled in. This does not have to crowd the answer. The value comes first; the record sits one step away, behind a “view sources” control, for anyone who wants to look. ## What this means for agents and answer engines For people building agents, the lesson is practical. Assume most readers will never open your citations, and design so that the answer can be checked anyway: keep each value tied to its record, keep the record’s date, and say plainly what was not established. That is what we are building IKAG to return for the specialist data it reaches. IKAG is opening to early users and data partners first. ## What the study does not tell us The numbers come from one month, one search engine and US adults who agreed to share their browsing. They describe clicks, not satisfaction or trust: a person who does not click may have got what they needed, or may have given up. Pew’s figures are a strong signal, not the last word, and AI search has changed since they were collected. ## Questions ### What did the Pew study measure? The online activity of 900 US adults who agreed to share their browsing data in March 2025, including 68,879 unique Google searches. Pew compared visits to results pages that showed an AI summary with those that did not. ### Does this mean people trust AI summaries? The study measured clicks, not trust. Fewer clicks can mean the summary answered the question, that people trusted it, or that they gave up. It does mean fewer people saw the original source. ### How often did AI summaries appear? About 18% of the Google searches in the study produced an AI summary, according to Pew. ### What should an answer carry if people will not click? The record behind each value: which source stated it, the record itself, when the source last attested it, and what it did not establish. A link alone asks the reader to do the checking. ## Sources - [Pew Research Center: Google users are less likely to click on links when an AI summary appears in the results (22 July 2025)](https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/) - [Columbia Journalism Review, Tow Center: AI Search Has a Citation Problem (6 March 2025)](https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php) --- # How to look up a company’s LEI with an AI agent > What a Legal Entity Identifier is, how an AI agent should query the GLEIF register, and why “no match” or “lapsed” never means a company is gone. - URL: https://ikag.dev/blog/how-to-look-up-a-company-lei-with-an-ai-agent - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-15 - Updated: 2026-10-15 - Reading time: 4 min - Publisher: IKAG ## Overview An LEI is a 20-character code that identifies a legal entity under ISO 17442, and GLEIF publishes the whole register free of charge. An agent should look companies up by identifier rather than by name, read the registration status separately from the entity status, and report a missing or lapsed record as exactly that, never as proof that a company does not exist. ## Key points - The Legal Entity Identifier (LEI) is a 20-character code defined by ISO 17442. GLEIF publishes the global register free of charge under the CC0 licence. - More than 3.1 million LEIs were active at the end of June 2026, according to GLEIF, and 92,000 organisations obtained one in that quarter alone. - Look up by identifier wherever you can. A name returns candidates: in our own test, the name of one exchange-traded fund matched three entities in the register. - A lapsed LEI means the registration was not renewed. On its own it says nothing about whether the company still operates, and an agent should say exactly that. A Legal Entity Identifier (LEI) is the closest thing the world has to a passport number for companies. If an AI agent needs to know exactly which legal entity it is dealing with, the LEI register is usually the first place to ask, and it is open to anyone. This guide covers what the code means, how an agent should query it, and the three mistakes that turn a correct lookup into a wrong answer. ## What is an LEI? An LEI is a 20-character code that identifies one legal entity. GLEIF, the Global Legal Entity Identifier Foundation, describes it as a unique alphanumeric code under ISO 17442, linked to reference information that answers “who is who” and “who owns whom”. The first eighteen characters identify the entity and the last two are check digits, so a mistyped code can be caught before anyone looks it up. Each record carries the entity’s legal name, its registered and headquarters addresses, its legal form, the jurisdiction it was formed in, and, for almost every entity, information about its direct and ultimate parents. The data is published under the CC0 licence, and GLEIF calls the LEI “a public good, available free of charge to all users”. ## How many organisations have one? More than 3.1 million LEIs were active at the end of the second quarter of 2026, according to GLEIF’s quarterly figures, with 92,000 issued in the quarter and growth of 3.0 per cent. GLEIF also reported that 87.9 per cent of LEIs were fully corroborated, and that 99 per cent of the LEI population reported parent information. Growth is uneven, and it follows regulation. India grew fastest after its central bank made the LEI compulsory for participants in the markets it regulates, and GLEIF attributes growth in Latvia and Lithuania to the continuing application of the EU’s Digital Operational Resilience Act. **Chart: Fastest-growing LEI jurisdictions, second quarter of 2026** | Item | Value | Note | | --- | --- | --- | | India | 7.1% | After the RBI made the LEI compulsory in its markets | | Latvia | 5.2% | Attributed to the EU’s DORA | | Brazil | 5.1% | | | Lithuania | 4.7% | Attributed to the EU’s DORA | | United Arab Emirates | 4.6% | | | Global average | 3.0% | | Quarterly growth in the number of LEIs, by jurisdiction, with the global rate for comparison. Source: [GLEIF: The LEI in numbers, Q2 2026 (23 July 2026)](https://www.gleif.org/en/newsroom/blog/the-lei-in-numbers-q2-2026-signals-growing-regional-momentum-for-the-lei). ## How should an agent query the register? By identifier, whenever it has one. A request for a single LEI returns one record, and that record either exists or it does not. GLEIF’s public API returns a record at a predictable address, for example `https://api.gleif.org/api/v1/lei-records/HWUPKR0MPOU8FGXBT394` for Apple Inc. In our own work, IKAG’s first capability reads it without an API key. GLEIF also publishes two MCP servers of its own, in beta: one searches the LEI register and the other answers questions about LEI statistics. GLEIF says they work with assistants such as ChatGPT, Claude and Mistral, and that a call sends only the parameters needed for the lookup. If you want your own assistant to look up LEIs today, that is the most direct route. By name only when it has to, and then as a search, not an answer. In a test of IKAG’s first capability on captured records, the name “SPDR S&P 500 ETF Trust” matched three entities in the register, one of them inactive. The right response to a name is the list of candidates and a question back to the user, not a guess. Three habits keep an agent fast and fair to the source: - **Cache by LEI.** Records change rarely; the record says when it was last updated, so store that date with the value. - **Pace requests.** Public registries limit how quickly any one client may call them. Treat a refusal as a signal to slow down, not as “no record”. - **Keep the record’s own dates.** A value fetched today can come from a record last updated months ago. Report the record’s date, not the fetch time. ## Registration status is not entity status An LEI record has two statuses, and agents often confuse them. | Field | What it describes | Common values | | ------------------- | ---------------------------------------- | ------------------------------- | | Registration status | The LEI registration itself | ISSUED, LAPSED, RETIRED, MERGED | | Entity status | Whether the entity is reported as active | ACTIVE, INACTIVE | A registration is **lapsed** when it has not been renewed and has passed any grace period for renewal. That is a fact about paperwork. GLEIF has been explicit that a lapsed LEI does not mean the associated legal entity is inactive. An agent that reads LAPSED and answers “this company no longer exists” has invented a fact. ## No match is not “no company” A search that finds nothing tells you about the search, not the world. A company may have no LEI because it never needed one, because it is registered under a different name, or because the identifier you searched with belongs to another system. When IKAG’s first capability asked GLEIF about Apple by its SEC identifier, the register returned no match: Apple’s LEI record does not name that SEC number. The company plainly exists; that record simply was not the place that links those two identifiers. The honest answer is “not established from this source”, with the reason. ## Linking an LEI to other identifiers Agents often hold more than one identifier for the same company. Join them only when a primary record names both. Wikidata stores LEIs under property P1278 and is useful for finding a candidate, but community data can agree with a value without proving it. In our tests, an LEI and an SEC CIK that no primary record linked were kept apart, and the result said why. ## What an agent should return A good LEI lookup returns more than a name. It returns: 1. **The value**, for example the legal name. 2. **The record it came from**: the LEI, the source, and a link to the record. 3. **The record’s own date**, when the source states one. 4. **The statuses**, registration and entity, reported separately. 5. **What was not established**, with the reason: no match, not in the record, or several candidates. That is the difference between an answer and a checkable answer. The second takes a few more fields, and it is the only kind you can rely on. ## Questions ### Is the GLEIF API free? GLEIF says LEI data is available free of charge to all users and publishes it under the CC0 licence. Public registries still limit how fast any one client may call them, so cache results and pace requests. ### How many characters is an LEI? Twenty. The first eighteen are letters and digits, and the last two are check digits. The format is defined by ISO 17442. ### Does a lapsed LEI mean the company has closed? No. Lapsed means the registration was not renewed in time. GLEIF has noted that a lapsed LEI does not mean the associated legal entity is inactive. Check the entity status and other records before saying anything about the company itself. ### Can an agent find an LEI from a company name? It can search by name, but a name is a lead, not an identity. Several entities can share a name, so an agent should return the candidates and ask which one is meant, or match on an identifier such as a registration number. ### Is an LEI the same as an SEC CIK? No. An LEI comes from the global LEI system and identifies a legal entity anywhere in the world. A CIK is a number the US Securities and Exchange Commission gives to filers. Treat a link between them as confirmed only when a primary record names both. ## Sources - [GLEIF: Introducing the Legal Entity Identifier](https://www.gleif.org/en/about-lei/introducing-the-legal-entity-identifier-lei) - [GLEIF: The LEI in numbers, Q2 2026 (23 July 2026)](https://www.gleif.org/en/newsroom/blog/the-lei-in-numbers-q2-2026-signals-growing-regional-momentum-for-the-lei) - [GLEIF: LEI data terms of use (CC0)](https://www.gleif.org/en/meta/lei-data-terms-of-use) - [GLEIF: How Banco de Portugal strengthens LEI data quality at scale](https://www.gleif.org/en/newsroom/blog/how-banco-de-portugal-strengthens-lei-data-quality-at-scale) - [GLEIF: Master Agreement appendix 07, payments (status definitions)](https://www.gleif.org/organizational-identity/become-an-lei-issuer/required-documents/2025-01-02_ma-appendix-07-payments_v1.5final.pdf) - [GLEIF: Use GLEIF data in your AI assistant (MCP connectors, beta)](https://ai.gleif.org/connect-ai) - [GLEIF API: LEI record HWUPKR0MPOU8FGXBT394](https://api.gleif.org/api/v1/lei-records/HWUPKR0MPOU8FGXBT394) - [Wikidata: property P1278, Legal Entity Identifier](https://www.wikidata.org/wiki/Property:P1278) --- # LEI, CIK or Wikidata ID: which company identifier? > A practical guide for agents and developers: what LEIs, SEC CIKs and Wikidata IDs identify, who issues them, how to resolve each, and when to link them. - URL: https://ikag.dev/blog/lei-cik-wikidata-company-identifiers-for-ai-agents - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-15 - Updated: 2026-10-15 - Reading time: 3 min - Publisher: IKAG ## Overview Use the LEI to identify a legal entity anywhere in the world, the SEC CIK to find a filer’s records in EDGAR, and a Wikidata ID to reach a community-maintained description that can corroborate a value but not prove it. Link two identifiers only when a primary record names both of them. ## Key points - An LEI is a 20-character ISO 17442 code from the global LEI system; a CIK is a 10-digit number the US SEC assigns to filers; a Wikidata ID (QID) labels an item in a community-edited knowledge base. - Each answers a different question, so a good agent keeps them apart and records which one each value came from. - Wikidata stores LEIs under property P1278 and CIKs under P5531, which makes it a useful lead. Community data should corroborate a value, not prove it. - Join an LEI and a CIK only when a primary record names both. In our tests, Apple’s LEI record and SEC record did not name each other, so they were kept apart. Ask an AI agent about a company and it will usually reach for a name. Names are where identity goes wrong: they change, they repeat, and they are spelt differently in every system. Identifiers fix that, but only if the agent knows which identifier answers which question. Here are the three an agent meets most often when it researches companies, and the rule for linking them. ## The three identifiers at a glance | Identifier | Issued by | Format | What it identifies | Access | | ----------------- | ------------------------------------------- | ---------------------------- | ------------------------------------------- | ------------------------------------- | | LEI | The global LEI system, overseen by GLEIF | 20 characters, ISO 17442 | A legal entity, anywhere in the world | Free; published under CC0 | | SEC CIK | The US Securities and Exchange Commission | 10 digits with leading zeros | A person or company that files with the SEC | Free; no key; fair-access rate limits | | Wikidata ID (QID) | Wikidata, a community-edited knowledge base | “Q” followed by digits | An item, which may describe a company | Free; structured data under CC0 | They overlap, but they are not interchangeable. A company can have an LEI and no CIK (most companies outside the US securities markets), a CIK and no LEI, or both. A Wikidata item can exist for any of them, and its claims are only as good as the edits behind them. ## When to use which - **“Which legal entity is this?”** Start with the LEI. Its record gives the legal name, the jurisdiction the entity was formed in, its legal form and, for nearly every entity, its parents. - **“What has this company filed?”** Use the CIK. The SEC’s data APIs need no authentication or key, update throughout the day as filings are disseminated, and address companies by the 10-digit CIK, for example `https://data.sec.gov/submissions/CIK0000320193.json`. The SEC asks automated clients to declare a user agent and to stay under 10 requests a second. - **“What else is known, and does anyone disagree?”** Use the Wikidata item. It is a fast way to find candidate identifiers: it stores LEIs under property P1278 and CIKs under property P5531. Treat what it says as corroboration, not proof. ## Can you link them? Only when a primary record names both. This is the rule that separates a careful agent from a confident one. In IKAG’s tests on captured public records, Apple’s LEI record and its SEC record did not name each other. Wikidata linked them, but Wikidata is community data, so the capability kept the two identifiers apart and said why. The SPDR S&P 500 ETF Trust was different: its LEI record names its CIK, so the two could be joined on the strength of a primary record. | Situation | Join them? | Why | | --------------------------------------- | ----------- | -------------------------------------- | | A primary record names both identifiers | Yes | The issuer of the record made the link | | Only a community source links them | No, not yet | A lead to confirm, not a fact | | The names match | No | Different entities often share names | ## Pitfalls that produce wrong answers 1. **Unpadded CIKs.** “320193” and “0000320193” are the same CIK, but the SEC’s APIs expect the ten-digit form. 2. **Names as identifiers.** A name search returns candidates. In one of our tests, a fund’s name matched three entities in the LEI register, one of them inactive. 3. **Lapsed is not closed.** A lapsed LEI means the registration was not renewed, not that the company stopped operating. 4. **Silent conflicts.** When sources disagree, the conflict is the answer. Report both values and their records rather than picking one quietly. ## What an agent should return For each identifier it uses, an agent should say where it came from: the issuing source, the record, and whether a primary record or only a community source supports any link to another identifier. That small amount of bookkeeping is what lets a reader, or another agent, check the answer in seconds instead of starting again. ## Questions ### Which identifier should an agent prefer? The one issued by the authority for the question. For “which legal entity is this?”, the LEI. For “what has this company filed with the SEC?”, the CIK. Wikidata is best used to find candidates and to check agreement. ### Why does a CIK sometimes have leading zeros? The SEC’s data APIs address companies by a 10-digit CIK with leading zeros, for example 0000320193 for Apple Inc. Pad shorter numbers to ten digits before you build a request. ### Is Wikidata reliable enough to link identifiers? It is a good lead and often right, but anyone can edit it. Use it to find a candidate link, then confirm the link in a primary record before you treat two identifiers as the same company. ### Can one company have several LEIs? A legal entity should have one LEI, but a corporate group has one per legal entity. A parent, its subsidiaries and its funds each have their own, linked through parent relationships in the LEI data. ## Sources - [GLEIF: Introducing the Legal Entity Identifier](https://www.gleif.org/en/about-lei/introducing-the-legal-entity-identifier-lei) - [SEC: EDGAR application programming interfaces](https://www.sec.gov/search-filings/edgar-application-programming-interfaces) - [SEC: Accessing EDGAR data (fair access)](https://www.sec.gov/search-filings/edgar-search-assistance/accessing-edgar-data) - [Wikidata: property P1278, Legal Entity Identifier](https://www.wikidata.org/wiki/Property:P1278) - [Wikidata: property P5531, Central Index Key](https://www.wikidata.org/wiki/Property:P5531) - [Wikidata: licensing (structured data under CC0)](https://www.wikidata.org/wiki/Wikidata:Licensing) - [GLEIF: LEI data terms of use (CC0)](https://www.gleif.org/en/meta/lei-data-terms-of-use) --- # Tool routers and data networks: what agents need > Two ways to connect AI agents to outside data: a router that offers many paid tools behind one key, and a data network that returns the record behind each value. - URL: https://ikag.dev/blog/tool-routers-and-data-networks-for-ai-agents - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-15 - Updated: 2026-10-15 - Reading time: 4 min - Publisher: IKAG ## Overview A tool router gives an agent one key and a single bill for many third-party tools, and leaves the choice of tool, and the reading of its output, to the agent. A data network starts from the data you need: it chooses sources by a policy, reconciles what they say, and returns each value with its record and anything it could not establish. They solve different problems, and many agents will use both. ## Key points - Tool routers put many APIs behind one connection, one key and per-call billing. The agent picks the tool and interprets what comes back. - A data network takes a request for data, chooses sources by policy, reconciles them, and returns values with their records and their unknowns. - Routers are about access and payment; data networks are about evidence. Neither replaces web search, and MCP can carry either. - IKAG is building a data network for AI agents, opening to early users and data partners first. AI agents need two things from the outside world: the ability to act, and facts to act on. Over the past year a crowd of services has grown up to connect agents to third-party APIs, and they tend to describe themselves in the same words. Underneath, there are two different ideas. One is a tool router. The other is a data network. Knowing which one you need saves a lot of glue code and a lot of wrong answers. ## Two ways to give an agent outside data Both put one connection between an agent and many providers. The difference is what they promise about the result. - A **tool router** promises access. It lists many tools, handles keys and billing, and passes each call through. What comes back is whatever the tool returns. - A **data network** promises evidence. It accepts a request for specific data, decides which sources to ask, and returns each value tied to the record that states it, with the gaps named. ## What a tool router does A router is a catalogue with a single front door. An agent browses or searches the catalogue, picks a tool, and calls it through the router, usually paying per call from a single prepaid balance instead of signing up with each provider. That is genuinely useful. One key replaces dozens, billing is in one place, and an agent can try a new tool without new paperwork. The trade-off is that the router leaves the hard questions to the agent: which tool is the right source for this question, whether its output is current, and whether two tools that disagree are describing the same thing. ## What a data network does A data network starts from the data, not the tool. The request says what is needed, such as a company’s legal name, its identifiers and where it is incorporated, and optionally which sources are acceptable, how fresh the data must be and what it may cost. The network then: 1. **Chooses sources by policy**, for example primary registries first and community data only as support. 2. **Calls them within their terms**, pacing requests and respecting each source’s limits. 3. **Reconciles the answers**, keeping conflicts as conflicts rather than picking a winner silently. 4. **Returns each value with its record**, and lists what it could not establish, with the reason. ## Side by side | Question | Tool router | Data network | | --------------------------------- | ----------------------------- | -------------------------------------- | | What do you ask for? | A tool | The data you need | | Who chooses the source? | The agent | The network, by a visible policy | | What comes back? | The tool’s raw output | Values, each with the record behind it | | What happens to gaps? | Usually silent | Reported as unknown, with a reason | | What happens when sources differ? | The agent has to notice | Kept as a conflict, with both records | | Best for | Actions and broad tool access | Facts an answer has to rest on | ## Which one do you need? If your agent mostly acts, sending messages, generating media or calling a long tail of utilities, a router is the convenient choice. If your agent mostly answers, and the answers have to be checkable, you need something that returns records and unknowns, or you will end up building that layer yourself. Many agents will use both: a router for doing, a data network for knowing. ## Where MCP fits The Model Context Protocol is how many agents now connect to either kind of service. Anthropic introduced it on 25 November 2024 as an open standard for connecting AI tools to data sources. On 9 December 2025 Anthropic donated it to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI, and reported more than 10,000 active public MCP servers and more than 97 million monthly SDK downloads across Python and TypeScript. MCP standardises the connection. It does not decide which source is right for a question, and it does not make a value checkable. Those are properties of the service behind the connection, which is why the router or network question matters even when everything speaks MCP. ## Where IKAG fits IKAG is building a data network for AI agents: one connection through which the AI people already use can reach specialist data sources, with each value tied to the record it came from where the source supports it. It is opening to early users and data partners first. Its first capability covers company legal identity from official registries, and you can [inspect real output from it](https://ikag.dev/agents). ## Questions ### Is a tool router the same as an MCP server? No. MCP is a protocol for connecting AI applications to tools and data. A router can offer its catalogue through an MCP server, and so can a data network or a single API. The protocol is the plug, not what sits behind it. ### Does a data network replace web search? No. Web search finds pages about a question. A data network goes to the sources that hold the record itself, such as official registries, and is the better fit when you need a specific, checkable value. ### Which should I use for company research? If you need many different tools and are happy to interpret their output yourself, a router is convenient. If you need values you can check, such as a company’s legal name and where it is incorporated, look for something that returns the record behind each value and reports what it could not find. ### Can one agent use both? Yes. An agent might use a router for actions such as sending a message or generating an image, and a data network when an answer has to rest on records. ## Sources - [Anthropic: Introducing the Model Context Protocol (25 November 2024)](https://www.anthropic.com/news/model-context-protocol) - [Anthropic: Donating the Model Context Protocol and establishing the Agentic AI Foundation (9 December 2025)](https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation) - [Model Context Protocol: introduction](https://modelcontextprotocol.io/docs/getting-started/intro) --- # What should the source record behind an AI answer contain? > Seven parts that make an AI answer checkable, from the source and record ID to attestation dates and unknowns, shown on seven real company lookups. - URL: https://ikag.dev/blog/what-the-source-record-behind-an-ai-answer-contains - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-15 - Updated: 2026-10-15 - Reading time: 4 min - Publisher: IKAG ## Overview A source record is the evidence attached to a value: which source stated it, which record, a link to that record, when the source last attested it, how the value was matched, and what the record did not establish. A link to a web page is not enough. In seven real company lookups, the records established 19 of 63 requested fields; the other 44 came back as unknown, each with a reason. ## Key points - A useful source record names the source, the record and its identifier, a link to it, when the source attested the fact, how the value was matched, and what the record does not say. - Freshness belongs to the source, not the fetch: a value fetched today can describe a record last updated months ago. - In seven real lookups by IKAG’s first capability, 19 of 63 requested fields were established. The other 44 were reported as unknown with a reason, never filled in. - Unknown is an answer. An agent that says what it could not establish is easier to check than one that fills every field. Most AI answers come with a link, if they come with anything. A link tells you where the answer says it came from. It does not tell you which record stated the value, how old that record is, or what the record left out. For answers you need to rely on, that is the information that matters. This article sets out what a source record should contain and shows it on real lookups. ## What is a source record? A source record is the evidence attached to a single value in an answer. If an assistant tells you a company is incorporated in California, the source record says which source stated that, in which record, when the source last attested it, and how confident the match was. It is attached to the value, not to the answer as a whole, because different values in one answer often come from different places. The idea is not new. Provenance standards such as the W3C’s PROV model have described where data comes from for years. What is new is that AI agents now assemble answers from many sources at once, and a reader has to be able to take any one value apart. ## The seven parts of a useful record | Part | The question it answers | Example from a real lookup | | --------------- | ---------------------------------------- | ------------------------------------------ | | Source | Who stated this? | SEC EDGAR, a primary registry | | Record | Which record, exactly? | The SEC company record for CIK 0000320193 | | Link | Where can I see it? | The record’s address at the source | | Attested | When did the source last state it? | Not stated by this source, so left blank | | Match | How was the entity identified? | By exact identifier, not by name | | Agreement | Does any other source agree or disagree? | Wikidata gives the same name | | Not established | What did the record not say? | Whether the company is active; its website | Two of these parts do most of the work. The record identifier lets anyone open the same record and check it. The “not established” list stops a reader from assuming the answer covered everything they asked. ## Freshness belongs to the source, not the fetch An agent that fetched a record five seconds ago will be tempted to call its values current. That is a mistake. The record may have been last updated months ago, and some sources do not say at all. In IKAG’s tests, GLEIF dated each record: Apple’s was last updated on 3 March 2026, however recently it was fetched. The SEC’s company facts carried no such date, so the values were never called current, even straight after fetching them. A good record keeps the source’s date where there is one and leaves the field empty where there is not. ## What real lookups return We asked IKAG’s first capability, company legal identity, for the same nine fields in seven different ways, on public records captured between 29 August and 7 October 2026. Across the seven requests, 19 of the 63 requested fields were established. The other 44 came back as unknown, each with a reason. **Chart: Fields established per request, out of 9, in seven real lookups** | Item | Value | Note | | --- | --- | --- | | SEC CIK and Wikidata ID | 4 of 9 | | | SEC CIK only | 4 of 9 | | | LEI only | 5 of 9 | | | Fund LEI only | 6 of 9 | | | LEI and SEC CIK | 0 of 9 | No primary record linked the two, so neither was assumed | | Wikidata ID only | 0 of 9 | Community data corroborates; it never proves | | Fund name only | 0 of 9 | Three entities matched the name; none was picked | Each request asked for the same nine fields of company legal identity. Red bars are requests where the honest answer was that nothing could be established. Source: [IKAG: seven requests to the first capability, on captured public records](https://ikag.dev/agents#runs). The three empty requests are not failures. In one, a request named an LEI and an SEC CIK that no primary record linked, so the capability did not assume they were the same company. In another, it had only a Wikidata ID, and community data can agree with a value but never prove one. In the third, a fund’s name matched three entities, and a name never selects a company. **Chart: How often each requested field was established, in the same seven lookups** | Item | Value | | --- | --- | | Legal name | 4 of 7 | | Identifiers | 4 of 7 | | Jurisdiction | 4 of 7 | | Entity status | 2 of 7 | | Registered address | 2 of 7 | | Latest filing | 2 of 7 | | Legal form | 1 of 7 | | Registration jurisdiction | 0 of 7 | | Official website | 0 of 7 | Some fields are rarely stated by these registries at all. Reporting them as unknown is more useful than filling them from elsewhere without a record. Source: [IKAG: seven requests to the first capability, on captured public records](https://ikag.dev/agents#runs). Some fields are simply rare in these registries. None of the seven lookups established a registration jurisdiction or an official website from the records consulted. The right answer for those fields is unknown, not a value borrowed from somewhere without a record. ## Unknown is an answer There are two kinds of unknown, and a record should say which one it is. - **Not observed.** The sources consulted were the right ones, and none of them stated the value. More sources might help. - **Insufficient evidence.** The request could not be tied to one entity with confidence, so no value was attached at all. More identifiers might help. An agent that reports unknowns this way gives you something to act on. An agent that fills every field gives you something to doubt. ## How to show it without burying the answer A record does not have to clutter the answer. The pattern that works is progressive: the value first, a small mark that it has a source, and the full record one step away, behind a “view sources” control. The unknowns sit under their own “what’s still unknown” control, so they are visible to anyone who looks and invisible to anyone who does not need them. The assistant decides how much of this to show. The source record’s job is to exist, complete, so that the assistant, or you, can. ## Questions ### Is a link to a web page a source record? Not on its own. A link shows where an answer says it came from. A source record also names the record, the date the source attested the fact, and what the record does not establish, so the claim can be checked without guessing. ### What is the difference between attested and retrieved? Retrieved is when a system fetched the data. Attested is when the source itself stated or last updated the fact. Only the second tells you how current the fact is. ### Should an agent fill a missing field from another source? Only if that source is suitable for the field and the value comes back with its own record. Otherwise the honest answer is unknown, with the reason. ### Does a source record make an answer correct? No. Records can be wrong or out of date. A source record makes an answer checkable: you can see what stated it and when, and decide whether to rely on it. ## Sources - [IKAG for builders: seven requests to the first capability](https://ikag.dev/agents#runs) - [W3C PROV-O: The PROV Ontology (30 April 2013)](https://www.w3.org/TR/prov-o/) - [GLEIF API: LEI record HWUPKR0MPOU8FGXBT394](https://api.gleif.org/api/v1/lei-records/HWUPKR0MPOU8FGXBT394) - [SEC: EDGAR application programming interfaces](https://www.sec.gov/search-filings/edgar-application-programming-interfaces) - [Wikidata: item Q312](https://www.wikidata.org/wiki/Q312) --- # How AI agents get live data and tools, and how to choose > Search, page extraction, specialist data APIs, generation models and MCP: how an AI agent reaches the outside world, with a worked example on real records. - URL: https://ikag.dev/blog/how-ai-agents-get-live-data - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-10 - Updated: 2026-10-10 - Reading time: 5 min - Publisher: IKAG ## Overview Search, page extraction, specialist data APIs, generation models and MCP: how an AI agent reaches the outside world, with a worked example on real records. ## Key points - An AI agent reaches outside information in four ways: search, reading web pages, specialist data APIs and your own systems. Models that make images, video or voice are a fifth kind of tool, but what they make is new content, not evidence. - Search finds pages about a question; a specialist source such as an official registry holds the record itself. - MCP is an open standard for connecting AI applications to tools and data. It standardises the connection; it does not choose the right source for you. - Choose by freshness, authority, coverage, cost and terms, and prefer tools that return the record behind each value. A language model on its own knows what it learned in training. To answer "what are people saying about our new menu this week?", "which legal entity is behind this fund?" or "what did this company file last month?", an agent has to reach something outside itself. ## The ways an agent reaches the outside world Four of them bring back information: 1. **Search.** A search API returns ranked results and snippets for a query. It is the fastest way to find out what exists and what is new. 2. **Reading web pages.** Page extraction and crawling tools turn pages into clean text, Markdown or structured data an agent can read. Use them when you know where the answer lives, such as a help centre or a pricing page. 3. **Specialist data APIs.** Official registries, maps and business listings, social media data, company and jobs data. These return structured facts, often more reliably than reading pages, from the organisation that keeps the record. 4. **Your own systems.** Internal databases, documents and tools, connected the same way. A fifth kind of tool makes things: **image, video and voice models**. What they return is new content, not information about the world. It can be exactly what you asked for, but it is not evidence, and the questions to ask are about quality, rights and cost rather than sources. Most useful requests combine several. A shortlist of local businesses might use listings to find them, reviews to rank them and their own websites to check opening hours. ## How agents connect to them Agents call outside systems through tools: functions the model can decide to use, with a name, a description and inputs. One open way to offer tools is the Model Context Protocol (MCP). In its own words, MCP is "an open-source standard for connecting AI applications to external systems." Anthropic introduced it in November 2024. Several assistants, including Claude and ChatGPT, can connect to MCP servers; what is available depends on the product, plan and settings. IKAG reaches assistants through its own MCP server. MCP standardises the connection. It does not decide which source is right for a request. If you connect ten MCP servers, the model chooses between ten sets of tools from their descriptions, and every server brings its own key, pricing and terms. ## A worked example: one question, three sources Take "which legal entity is behind Apple, and where is it incorporated?" Search would find many pages about Apple. A registry holds the record itself, so this is how the question goes through public registries, using records captured on 29 August 2026 (SEC EDGAR) and 7 October 2026 (GLEIF and Wikidata). - **A name alone is not enough.** Names are not unique. Searching the global LEI register for the name "SPDR S&P 500 ETF Trust", for example, returns three different entities, one of them inactive. A careful tool lists them and asks which one you mean. - **An identifier is.** With Apple's SEC identifier, CIK 0000320193, the [SEC EDGAR record](https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0000320193) gives the legal name, Apple Inc., and California as the state of incorporation. - **Community data can agree, not prove.** Wikidata's item for Apple gives the same name. That is useful support, but Wikidata is edited by its community, so it never decides the answer on its own. - **Some things stay unknown.** The SEC record does not say whether the company is active, and it is not dated, so you cannot call its facts current. The GLEIF register has no LEI registered against that CIK, so an LEI cannot be attached from these records alone. | Field | Value | From | | --------------- | ------------------------- | -------------------------- | | Legal name | Apple Inc. | SEC EDGAR; Wikidata agrees | | Incorporated in | California, United States | SEC EDGAR | | Active? | Not in these records | None | | LEI | Not linked by a record | None | The answer is narrower than "everything about Apple", and every part of it can be checked. ## How to choose a source | Kind of source | Best for | Watch for | | ----------------------- | --------------------------------------------------- | --------------------------------------------- | | Search | Finding things, news, broad questions | Snippets are not answers; check the page | | Reading web pages | Known pages, documentation, prices | Site terms and robots.txt; pages change shape | | Specialist data APIs | Structured facts, registries, listings, social data | Coverage gaps; pricing by plan, usage or both | | Your own systems | Private knowledge | Access control | | Models that make things | Images, video, voice | Quality varies by model; usage rights | Five questions decide most choices: - **Freshness.** How recent does the answer need to be, and does the source say when it last updated the fact? - **Authority.** Is there a primary source, such as an official registry? - **Coverage.** Does this source cover the place, language or platform you need? - **Cost.** What does a request cost, and how is it billed? - **Terms.** May you use the data this way? Respect robots.txt (standardised as RFC 9309 in 2022), terms of service and data protection law. ## Why provenance matters When an agent combines several sources, the answer is only as trustworthy as your ability to check it. Prefer sources and tools that return the record behind each value and when the source last updated it, and that say what they could not establish. Then a person can verify the answer, and an agent can weigh sources instead of guessing. ## Where IKAG fits IKAG is building one connection between AI assistants and agents and specialist sources like these. It is opening to early users and data partners first. Its first capability, company legal identity from the GLEIF and SEC EDGAR registries with Wikidata as support, run offline on captured records, behaves as the example above does. Web pages, listings, social media data and other sources come next. You can inspect [real output from the first capability](https://ikag.dev/agents). ## Questions ### Can AI agents browse the web? Many can, through search and page extraction tools connected to them. What they can reach depends on those tools and on each site’s rules. ### Is web scraping allowed? It depends on the site’s terms, the kind of data (personal data has extra legal protection) and the law where you operate. Respect robots.txt and terms of service, and avoid personal data without a lawful basis. This is not legal advice. ### What is MCP? The Model Context Protocol is an open-source standard for connecting AI applications to external systems such as data sources, tools and workflows. Anthropic introduced it in November 2024. ### Is an image or video made by a model a source? No. A generated image, video or voice clip is new content. It can be useful, but it is not evidence that anything is true, and it needs its own checks for quality and rights. ## Sources - [What is the Model Context Protocol (MCP)?](https://modelcontextprotocol.io/docs/getting-started/intro) - [Anthropic: Introducing the Model Context Protocol (25 November 2024)](https://www.anthropic.com/news/model-context-protocol) - [OpenAI: building apps and MCP servers for ChatGPT](https://developers.openai.com/chatgpt) - [RFC 9309: Robots Exclusion Protocol (September 2022)](https://www.rfc-editor.org/rfc/rfc9309.html) - [GLEIF API: LEI records](https://api.gleif.org/api/v1/lei-records/HWUPKR0MPOU8FGXBT394) - [SEC EDGAR: company filings for CIK 0000320193](https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0000320193) --- # How to check where an AI answer came from > A five-step check for any AI answer, worked through on real public records, and what good provenance looks like when you need to rely on a fact. - URL: https://ikag.dev/blog/how-to-check-ai-answers - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-10 - Updated: 2026-10-10 - Reading time: 4 min - Publisher: IKAG ## Overview A five-step check for any AI answer, worked through on real public records, and what good provenance looks like when you need to rely on a fact. ## Key points - An AI answer is a claim. A link under it is not proof that a source supports it. - In a March 2025 test by the Tow Center at Columbia Journalism Review, eight AI search tools asked to identify the source of quoted news excerpts got more than 60 percent of those queries wrong. That is one task, not an overall accuracy rate. - To check an answer: open the source, find the claim, prefer the primary source, check the date, look for agreement, and notice what the sources do not say. - Good provenance names the record behind each value, when that record was last updated, and what could not be established. AI assistants are very good at sounding sure. That is not the same as being right, and a link under an answer is not the same as a source that supports it. ## Confident is not the same as correct In March 2025, researchers at the Tow Center for Digital Journalism at Columbia Journalism Review tested eight AI search tools on one task: given a direct excerpt from a news article, identify its headline, original publisher, publication date and web address. Across 1,600 queries, in their words, "collectively, they provided incorrect answers to more than 60 percent of queries." They also found tools citing syndicated copies instead of the original articles, and fabricating links. That figure describes one demanding task with the tools of early 2025. It is not a general accuracy rate for AI, and tools change. The lesson that lasts is simpler: an answer is a claim, and a claim is only as good as the source you can check it against. ## A five-step check for any AI answer 1. **Open the source.** Does the link work, and does it lead to the page the answer names? 2. **Find the claim.** Does the page actually say what the answer says? Search the page for the key fact or number. 3. **Prefer the primary source.** A company's registered name comes from a company registry; a quote comes from the publisher; a train time comes from the operator. A summary of a summary is weaker. 4. **Check the date.** Is the source recent enough for the question? Opening hours, prices and news change. Note when the source says the fact was last updated, not only when you looked. 5. **Look for agreement, and for gaps.** Do two independent sources say the same thing? If they disagree, the answer should say so. And notice what the sources do not say: a gap is information too. ## A worked check, on real records Take a simple question: which legal entity is behind Apple, and where is it incorporated? An assistant might answer "Apple Inc., incorporated in California." Here is the check, on public records that anyone can open. 1. **Open the source.** The US Securities and Exchange Commission keeps a record for every company that files with it. Apple's is under Central Index Key (CIK) 0000320193 on [SEC EDGAR](https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0000320193). 2. **Find the claim.** The record names "Apple Inc." and gives California as its state of incorporation. The answer holds up. 3. **Prefer the primary source.** The SEC record is a primary source for a US filer. [Wikidata's item Q312](https://www.wikidata.org/wiki/Q312) gives the same name, which is useful agreement, but it is community-edited, so it supports the answer rather than proving it. 4. **Check the date.** The SEC record does not say when its company facts were last confirmed, so you cannot call them current; you can only say they are what the record shows. The global LEI register is better here: [Apple's LEI record](https://api.gleif.org/api/v1/lei-records/HWUPKR0MPOU8FGXBT394) states it was last updated on 3 March 2026. 5. **Look for agreement, and for gaps.** The LEI record also names Apple Inc. in California, under a California registration number, but it does not list Apple's SEC CIK, and the SEC record does not list an LEI. Both describe "Apple Inc.", so it is very likely the same company, but no record in this check links the two identifiers. The honest summary is: name and place confirmed by a primary record; the link between the LEI and the CIK not established by these records. That last step is the one most answers skip. A careful answer says what it could not confirm. ## What good provenance looks like Provenance is the record of where an answer came from. The difference between weak and strong provenance is easy to see once you know what to look for: | Aspect | Weak provenance | Strong provenance | | ------------ | ---------------------------------- | -------------------------------------------------------- | | Sources | A few links under the whole answer | The record behind each value | | Freshness | Unknown | When the source last updated each value, if it says | | Disagreement | Hidden | Conflicts shown, and which source was preferred, and why | | Gaps | Filled in with a guess | Named as unknown, with the reason | | Coverage | Unclear what was checked | Sources consulted, used and skipped | When an answer comes back like this, a person can check it in seconds, and an agent can weigh sources instead of picking one at random. Even then, a record can be wrong or out of date: provenance tells you what to check, not that the answer is true. ## How IKAG approaches it IKAG is opening to early users and data partners first. Its first capability, company legal identity, is designed to answer the way the worked check above does: each value tied to the primary record that states it, Wikidata used only as supporting evidence, freshness taken from the source's own dates, and no link between an LEI and a CIK unless a primary record names both. You can inspect real output from it on the [developer page](https://ikag.dev/agents#runs). ## Questions ### Is a cited answer always right? No. A citation shows where an answer says it came from. You still need to check that the source exists, says what the answer claims, and is current enough for your question. ### What did the Tow Center study measure? In March 2025 it gave eight AI search tools direct excerpts from news articles and asked for each article’s headline, publisher, date and web address. Collectively they answered more than 60 percent of the 1,600 queries incorrectly. It measured that task only, not AI accuracy in general. ### What is a primary source? The organisation that creates or officially records the fact: a company registry for a company’s legal name, the publisher for an article, the operator for a timetable. ### If two records name the same company, are they the same company? Not necessarily. A shared name is a lead, not proof. Look for a record that names both identifiers, or treat the link as unconfirmed. ## Sources - [Columbia Journalism Review, Tow Center: AI Search Has a Citation Problem (6 March 2025)](https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php) - [SEC EDGAR: company filings for CIK 0000320193](https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0000320193) - [GLEIF API: LEI record HWUPKR0MPOU8FGXBT394](https://api.gleif.org/api/v1/lei-records/HWUPKR0MPOU8FGXBT394) - [Wikidata: item Q312](https://www.wikidata.org/wiki/Q312) --- # What to ask before you connect an agent to a data gateway > Eight questions for any agent data gateway, applied to IKAG in early access: provenance, failures, routing, cost, own keys, retention, terms. - URL: https://ikag.dev/blog/questions-before-you-connect-an-agent - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-10 - Updated: 2026-10-10 - Reading time: 4 min - Publisher: IKAG ## Overview Eight questions for any agent data gateway, applied to IKAG in early access: provenance, failures, routing, cost, own keys, retention, terms. ## Key points - A data gateway gives an AI agent one connection to many outside data and tool providers. Choosing one is a trust decision as much as a technical one. - Ask where each value came from, what happens when sources fail or disagree, how providers are chosen, and what a call costs before it runs. - Ask whether you can use your own keys, what is stored and for how long, whether its numbers are dated, and on what terms the data is used. - We apply the same eight questions to IKAG below. Some are answered by its first capability; several are not answered yet, and we say which. A data gateway gives an AI agent one connection to many outside providers: search, web pages, listings, company records and more. It can save you a stack of integrations. It also sits between your agent and everything it learns, so choosing one is a trust decision. For a single provider you rely on, a direct integration may still be simpler. These are the eight questions we would ask, and at the end we answer them for IKAG as it stands today. ## 1. Where did each value come from? Not "which sources did you use", but which source supplied this specific value. Ask to see a result. Good provenance names the record behind each value, when that source last updated it, and whether other sources agreed. A receipt that shows only what a call cost tells you about your bill, not about the answer. ## 2. What happens when a source fails or sources disagree? Every provider fails sometimes. Ask whether a failed call is marked as failed, whether it is retried with another provider, and whether you are charged for it. When two sources disagree, the result should say so and show which one was preferred and why, not quietly pick one. And when the sources do not settle a question, the result should say "unknown", not guess. ## 3. How is a provider chosen, and can I change it? Behind one connection, something decides which provider answers. Ask what that decision is based on: authority, price, speed, coverage, or several providers at once for comparison. You should be able to see the choice and set the policy, including excluding a provider or keeping to one you trust. ## 4. What will a call cost before it runs? Ask whether you can see the cost of a call before it runs, and set a ceiling so an agent stuck in a loop cannot run up a bill. Usage pricing only helps if the price is visible before you commit. ## 5. Can I use my own keys? If you already pay a provider, you may want to keep that plan and route calls through your own account. Ask how calls made with your own key are billed, by whom, and whether the gateway charges anything for its part. ## 6. What do you store, for how long, and who processes it? Requests can contain personal or commercial information. Ask what is logged, how long it is kept, where it is processed, and which subprocessors see it. If requests contain personal data, data protection law such as the UK GDPR or the EU GDPR may apply, depending on where you, your users and the gateway are; that is a question for the gateway and for your own legal advice. A security page with no answer to these questions is an answer. ## 7. Are your numbers consistent and dated? Catalogue sizes, provider counts and benchmark results are easy to inflate and quick to go stale. Ask for a dated figure and how it was counted, and check that the same number appears everywhere the gateway publishes it. If benchmarks are quoted, ask for the method and the raw data. ## 8. On what terms is the data used? Some data comes from official APIs, some from licensed feeds, and some from reading public pages. Ask which is which for the sources you rely on, whether the gateway respects robots.txt (the Robots Exclusion Protocol, RFC 9309) and site terms, and how it handles personal data found on public pages. ## The same questions, asked of IKAG IKAG is opening to early users and data partners first. Here is how it approaches each question; where something isn’t decided, we say so. | Question | How IKAG approaches it | | ------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ | | 1. Each value's source | Every value names the record it came from, and when that source attested it, if it says. [See real output](https://ikag.dev/agents#real-result). | | 2. Failures, conflicts | Conflicts are kept as conflicts, gaps are reported as unknown with a reason, and a rate-limited source is refused, not retried silently. Whether failed calls cost anything will be published with prices. | | 3. Choosing a source | Primary registries first, community data only as support, and a request can exclude sources. | | 4. Cost before a call | Runs are quoted before they execute and carry a cost ceiling. Prices will be published before IKAG charges anyone. | | 5. Your own keys | The provider bills you under your plan, and IKAG may charge separately for its part. | | 6. Storage and retention | We will publish what the service stores, for how long and who processes it before it opens. | | 7. Numbers | We publish none yet. We will only publish figures that are real, dated and counted the same way everywhere. | | 8. Terms of data use | The first capability uses public records: GLEIF and Wikidata publish under CC0, and SEC EDGAR is public US government data. Commercial sources depend on an agreement with each provider; none is announced. | If that is the kind of gateway you want, see [real output from IKAG’s first capability](https://ikag.dev/agents). ## Questions ### What is an agent data gateway? A service between an AI agent and many outside data and tool providers, so the agent connects once instead of integrating each provider. Accounts and billing vary: some providers still need your own account or key. ### Should failed calls cost anything? That depends on the gateway’s terms, so ask. At minimum, a failed call should be clearly marked as failed on your receipt, never presented as an answer. ### Does data protection law apply to what my agent sends? It can. If requests contain personal data, laws such as the UK GDPR or the EU GDPR may apply, depending on where you, your users and the gateway are. Ask the gateway, and take legal advice for your situation. This is not legal advice. ## Sources - [ICO: UK GDPR guidance and resources](https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/) - [Regulation (EU) 2016/679 (General Data Protection Regulation)](https://eur-lex.europa.eu/eli/reg/2016/679/oj) - [RFC 9309: Robots Exclusion Protocol](https://www.rfc-editor.org/rfc/rfc9309.html) --- # What is an intelligent data network for AI agents? > What IKAG, in early access, means by an intelligent data network for AI agents, and how it compares with single APIs, marketplaces and MCP servers. - URL: https://ikag.dev/blog/what-is-an-intelligent-data-network - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-10 - Updated: 2026-10-10 - Reading time: 3 min - Publisher: IKAG ## Overview What IKAG, in early access, means by an intelligent data network for AI agents, and how it compares with single APIs, marketplaces and MCP servers. ## Key points - An intelligent data network gives an AI assistant or agent one connection to many external data sources, chooses the source for each request by a policy you can see, and returns each value with the record it came from. - It is one way to connect agents to data. Single APIs, API marketplaces, tool gateways and MCP servers each suit different needs, and for one dependable provider a direct integration may be simpler. - IKAG is building one, opening to early users and data partners first; its first capability covers company legal identity from official registries. An **intelligent data network for AI agents** is a single connection that lets an AI assistant or agent reach many external data sources, chooses which one to use for each request by a policy you can see, and returns each value with the record it came from. It is our name for what IKAG, short for I Know A Guy, is building; it describes a design, not an established market category. ## The problem it is for Language models are good at reasoning over what they are given. They are not, on their own, good at knowing what a company's registered name is, what was filed this week, or which of three funds with similar names you mean. For that, an agent has to call something outside itself: a search engine, a web page, an official registry, a listings database, a social media data provider. Each of those providers has its own API, keys, pricing and terms. Connecting an agent to one is straightforward. Connecting it to many, keeping those connections working, and knowing which one to trust for which question is a project in itself. ## How it works A request moves through the network in four steps: 1. **The agent asks.** It sends a request through one connection, such as an MCP server or a REST API. 2. **The network works out what is needed.** "Which legal entity is behind Apple?" needs company identity data from a registry. 3. **It picks a source by policy.** For legal identity, primary registries come first and community data only supports them. Other policies can favour price, speed or coverage, and a request can exclude sources. 4. **It returns what the records establish.** Each value comes back with the record behind it, conflicts stay visible, and anything the records do not settle is reported as unknown. ## How it compares with what you may already use Each approach suits different needs. These are the trade-offs as we see them; individual products vary, so check the ones you are considering. - **A single API.** Simple and direct when one provider covers what you need. You build and maintain the integration, and you decide in code when to call it. - **An API marketplace.** One place to find and subscribe to many APIs. You still integrate and choose each one yourself. - **A tool gateway.** One key and one connection to many tools, often with per-call billing. How much each result says about where its values came from varies by gateway. - **An MCP server.** A standard way to offer tools to an AI application. The model chooses between the tools it is shown, from their descriptions. - **An intelligent data network.** One connection to many sources, a source chosen per request by a visible policy, and the record behind each value returned with it. What we think matters most is the last part. When each value comes back with its record, a person can check it and an agent can weigh it. When the records do not settle something, saying "unknown" is part of the answer. ## What to look for, whatever you choose - **Provenance.** Does each value say which record it came from, and when that source last updated it? - **Routing you can see.** Can you tell why a source was chosen, and change the policy? - **Clear costs.** Can you see what a call will cost before it runs, and who bills you for it? - **Your own keys.** Can you keep provider accounts you already pay for? - **Honest failure.** When sources disagree or none can answer, does it say so? ## What IKAG is building IKAG is opening to early users and data partners first. Its first capability, company legal identity, answers "which legal entity is this, and where is it registered?" from the GLEIF and SEC EDGAR registries, with Wikidata as supporting evidence. You can [inspect real output from it](https://ikag.dev/agents). IKAG is not affiliated with those organisations. When IKAG opens, the AI you use will reach it through the Model Context Protocol, a REST API or TypeScript and Python SDKs. Pricing is designed around usage, and no prices have been set. ## Questions ### Is an intelligent data network the same as an MCP server? No. An MCP server is one way to offer tools to an AI application, and usually exposes one service. A network can be reached through MCP, but behind that one connection it chooses between many sources for each request. ### Does it replace web search? No. Web search is one kind of source. For many questions a specialist source, such as an official registry, holds the record itself, and the network picks it when it is the better fit. ### Can I use IKAG now? IKAG is opening to early users and data partners first. ## Sources - [Model Context Protocol: introduction](https://modelcontextprotocol.io/docs/getting-started/intro) - [GLEIF: the Global Legal Entity Identifier Foundation](https://www.gleif.org/) - [SEC EDGAR full-text search and filings](https://www.sec.gov/edgar/search/) - [Wikidata](https://www.wikidata.org/) --- # What IKAG’s first capability does, and won’t do > Seven offline requests to IKAG’s first capability, in early access: what it answers, what it leaves unknown, and why a shared name never joins two records. - URL: https://ikag.dev/blog/what-our-first-capability-does - Author: Ronit Ahuja, Founder, IKAG (https://ikag.dev/authors/ronit-ahuja) - Published: 2026-10-10 - Updated: 2026-10-10 - Reading time: 3 min - Publisher: IKAG ## Overview Seven offline requests to IKAG’s first capability, in early access: what it answers, what it leaves unknown, and why a shared name never joins two records. ## Key points - IKAG’s first capability, company legal identity, answers who a company legally is from the GLEIF and SEC EDGAR registries, with Wikidata as supporting evidence. IKAG is opening to early users and data partners first. - Given an identifier, it returns each value with the record that states it. Given only a name, it lists the matching entities and picks none. - It joins an LEI and an SEC CIK only when a primary record names both. Apple’s two records do not link each other, so it keeps them apart; the SPDR S&P 500 ETF Trust’s LEI record names its CIK, so those are joined. - Anything the records do not establish, such as a company’s status in an SEC record, comes back as unknown with the reason. IKAG’s first capability is deliberately narrow. Given a company, it answers who that company legally is: its registered name, its identifiers, where it is incorporated, its status, its registered address and its latest SEC filing, each tied to the record that states it. It draws on two primary registries, GLEIF (the global register of Legal Entity Identifiers) and SEC EDGAR, with Wikidata as supporting evidence. IKAG is opening to early users and data partners first. Everything below comes from IKAG’s own implementation, run offline on public records captured on 7 October 2026 (GLEIF and Wikidata) and 29 August 2026 (SEC EDGAR). The [developer page](https://ikag.dev/agents#runs) shows the raw output. ## Seven requests, and what came back | Asked with | What IKAG returned | | --------------------------------- | ---------------------------------------------------------------------------------------- | | Apple’s SEC CIK and Wikidata item | Apple Inc., incorporated in California, from the SEC record; Wikidata agrees on the name | | Apple’s SEC CIK only | The same, without Wikidata’s agreement | | Apple’s LEI only | Name, place, active status and registered address from GLEIF, last updated 3 March 2026 | | Apple’s LEI and SEC CIK together | No identity: no primary record names both, so they were not joined | | Apple’s Wikidata item only | No identity: community data cannot establish one alone | | A fund’s LEI only | SPDR S&P 500 ETF Trust, a New York fund, active, with its SEC CIK linked by GLEIF itself | | A fund’s name only | Three matching entities, one inactive; none chosen | ## A name is never enough Ask for "SPDR S&P 500 ETF Trust" by name and the LEI register returns three entities whose names match: the trust itself in New York, a Massachusetts fund with a similar name that is now inactive, and a Delaware trust whose name contains the same words. A tool that picked the first match would usually be right, and sometimes badly wrong. IKAG returns the three candidates and leaves the choice to the person or agent who asked, with an identifier. ## A shared name does not join two records Apple has an LEI, HWUPKR0MPOU8FGXBT394, and an SEC identifier, CIK 0000320193. Both records name Apple Inc. in California. But the LEI record gives a California registration number and names no SEC CIK, and the SEC record lists no LEI. Wikidata claims both identifiers belong to Apple, but Wikidata is community-edited. So when both identifiers are given together, IKAG does not merge them on the strength of a shared name. It reports the identity as unresolved, with the reason "insufficient evidence". That is less convenient than a confident merge, and it is the honest answer from these records. ## When a record does link them The SPDR S&P 500 ETF Trust is different. Its LEI record, 549300NZAMSJ8FXPQQ63, is registered with SEC EDGAR itself, under CIK 0000884394, and GLEIF has validated that registration. A primary record names both identifiers, so IKAG joins them as one entity, and says the link came from GLEIF. ## Unknown is an answer The SEC record says where Apple is incorporated, but not whether the company is active, and it states no official website. GLEIF holds no filings. So, depending on which identifiers you ask with, some fields come back as unknown, each with a reason such as "not observed" or "insufficient evidence". IKAG never fills a gap with a plausible guess. ## Dates come from the source A fact is only as fresh as the source says it is. GLEIF dates each record: Apple’s was last updated on 3 March 2026, however recently IKAG fetched it. The SEC’s company facts carry no such date, so IKAG never calls them current, even straight after fetching them. The latest filing it reports, a Form 4 accepted on 27 August 2026, is the latest in the record it read, not a promise that nothing newer exists. ## How you reach it IKAG is opening to early users and data partners first. It is reached through the Model Context Protocol, a REST API and TypeScript and Python SDKs, and this site lists each one as it opens. You can [inspect the raw output](https://ikag.dev/agents). ## Questions ### Can I use this capability today? IKAG is opening to early users and data partners first, and the early-access list is how you get a place. ### Why not match companies by name? Names are not unique and change over time. A name search in the global LEI register for “SPDR S&P 500 ETF Trust” returns three entities. Choosing one silently would be a guess. ### Is Wikidata a reliable source? It is useful supporting evidence and a good way to find identifiers, but anyone can edit it. IKAG uses it to corroborate values and suggest links, never to prove a legal identity on its own. ## Sources - [GLEIF API: LEI record HWUPKR0MPOU8FGXBT394 (Apple Inc.)](https://api.gleif.org/api/v1/lei-records/HWUPKR0MPOU8FGXBT394) - [GLEIF API: LEI record 549300NZAMSJ8FXPQQ63 (SPDR S&P 500 ETF Trust)](https://api.gleif.org/api/v1/lei-records/549300NZAMSJ8FXPQQ63) - [SEC EDGAR: company filings for CIK 0000320193](https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=0000320193) - [Wikidata: item Q312](https://www.wikidata.org/wiki/Q312)