What is MCP? A plain guide to letting your AI do more

2026-09-27, Connect

MCP, the Model Context Protocol, is an open standard that lets an AI assistant such as Claude or ChatGPT connect to outside services and use them: read your calendar, search a company's documents, or place an order on your behalf. Without it, an assistant only knows what it learned in training and what you type. With it, the assistant gains tools, and it uses them in the middle of a normal conversation.

The project's own site describes MCP as a USB-C port for AI applications: one standard plug, so any assistant can connect to any service that supports it.

Who made MCP

Anthropic, the company behind Claude, introduced the Model Context Protocol in November 2024 as an open standard. Anyone can build with it, free of charge, and the specification is public.

It spread fast. Other AI companies added support, and in December 2025 Anthropic donated MCP to the Agentic AI Foundation, a new fund under the Linux Foundation co-founded with Block and OpenAI. That means no single company owns its future. Today MCP works in Claude, ChatGPT, Gemini, Microsoft Copilot, Visual Studio Code, Cursor and many other tools.

Why it matters

Before MCP, every pairing of an assistant and a service needed its own custom integration. A calendar company that wanted to work with three assistants had to build three different things, and most never did.

MCP turns that into one job on each side. A service builds one MCP server. An assistant supports MCP once. After that, they can work together, and so can every other pair that speaks the protocol. For you, that means your assistant can do far more than it could on its own, using services you choose.

The parts, in plain words

The server is the service's side. It is a program that offers a list of things the assistant can use. The most important are tools: actions such as "search the files", "create an event" or "place an order". A server can also offer resources, which are pieces of information to read, and prompts, which are ready-made instructions.

The client is the assistant's side. It lives inside the app you use, such as Claude, ChatGPT or a coding tool, connects to a server, asks what tools it offers, and passes those tools to the AI model. When the model decides a tool would help with your request, the client calls it and hands the answer back.

A connector is what the apps call a connected server. In Claude and ChatGPT you add a connector in the settings, often by pasting the server's address. The words differ between apps, and the idea is the same.

Local and remote servers. Some servers run on your own computer, for example one that lets a coding assistant read files in a folder. Others run on the internet at an address, and those are the ones you add to Claude or ChatGPT as connectors.

How sign-in works

A remote server usually needs to know who you are. MCP handles that with OAuth, the same standard behind "Sign in with Google" buttons across the web.

When you add a connector, your assistant opens the service's own sign-in page. You sign in there, on the service's site, and approve what the assistant may do. The assistant receives a token, a kind of key with limited rights. It never sees your password. You can take the key back at any time, usually from your account page on the service, and the assistant's access ends.

Staying safe with MCP

MCP gives an assistant real abilities, so a little care pays off. Claude's own guidance makes the same points.

• Connect only services you trust. A connector is a company acting for you. Check who runs it and where it keeps your data.
• Read what you allow. The sign-in step shows what the assistant may do. A good server asks for no more than it needs.
• Watch actions that cost money or send things. Keep an eye on what your assistant does with a new connector, especially at first.
• Be aware of hidden instructions. A server's text reaches the AI model, and a malicious one could try to steer it. Trusted sources are your best protection.
• Remove what you no longer use. Revoke old connectors in the service's account page and remove them from your assistant.
• Think about sensitive data. Before sharing health or financial details through a connector, read how the service stores and protects them.

A worked example: Care MCP

Care MCP is an MCP server that connects your assistant to certified personal trainers and nutrition coaches. It shows each part above in action.

1. Connect. You add the address https://caremcp.com/mcp as a connector in Claude or ChatGPT.
2. Sign in. Your assistant opens our sign-in. You sign in with Apple or Google and allow it. That is OAuth at work.
3. Tools appear. Your assistant now has a handful of tools: read the catalog, order a plan, an evaluation or a review of a plan, read your orders and documents, and see your credit.
4. You ask in plain words. "I want a certified trainer to write me an eight week plan." Your assistant knows which tool fits, asks you what the trainer needs, and sends the order.
5. A person does the work. A certified trainer writes the plan, usually within two days of taking the order. Your assistant reads it back to you.

The limits show good MCP design too. Your assistant cannot pay, change your account or see anyone else's orders, and you can revoke it at any time. You add credit yourself, and a document is charged only when it is delivered. How we keep briefs safe is on our security page, and the whole flow is explained in our guide to personal training via MCP.

For companies building AI products, the same service is reachable by API, which our article on adding a human coach to an AI fitness app describes.

Try your first connector

The easiest way to understand MCP is to add one connector and use it. Our step-by-step guide for Claude takes about two minutes, and the connect page has the steps for ChatGPT and other assistants.