What is MCP? A plain-English guide for AI users

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 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 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, 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 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 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 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 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 are for personal accounts in the US. On Grok, 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. The newer agents connect too, each in its own way: see ChatGPT Dots, Grok Bot and 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.
• 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 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 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 and Meta 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.

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 has the steps for each app. IKAG is opening to early users and data partners first.