Short answer
MCP, the Model Context Protocol, is an open standard that lets AI applications such as Claude and ChatGPT connect to outside tools and data. A business runs an MCP server that lists what it can do, such as searching products, and an AI assistant calls those tools on a person's behalf. Anthropic introduced MCP in November 2024, and it is now hosted by the Agentic AI Foundation under the Linux Foundation.
Key takeaways
- MCP lets AI assistants call tools and read data from outside services.
- An MCP server lists tools, resources and prompts that assistants can use.
- The July 2026 spec made MCP stateless and added server/discover.
What is MCP?
MCP, the Model Context Protocol, is an open standard for connecting AI applications to outside systems: data, tools and workflows. The MCP project defines it as "an open-source standard for connecting AI applications to external systems."
The project's own comparison is the easiest way to picture it: "Think of MCP like a USB-C port for AI applications." A USB-C port lets one laptop connect to many devices without a special cable for each. MCP lets one AI assistant connect to many services without custom code for each.
Anthropic introduced MCP on 25 November 2024, calling it "a new standard for connecting AI assistants to the systems where data lives." It has since spread well beyond Anthropic. The MCP documentation lists support in Claude, ChatGPT, Visual Studio Code and Cursor, among others. For a one-line definition, see our MCP glossary entry.
How does MCP work?
An AI application connects to an MCP server, asks what the server offers, and then calls what it needs while helping a person.
The MCP architecture has three parts:
- Host. The AI application a person uses, such as Claude or ChatGPT.
- Client. A connector inside the host. The host creates one client for each server it connects to.
- Server. A program that offers tools and data to the assistant. This is the part a business would run.
A server can offer three kinds of things, which MCP calls primitives:
- Tools, "Executable functions that AI applications can invoke to perform actions," such as searching a catalog or checking an order.
- Resources, data sources that give the assistant context, such as a product list or a policy page.
- Prompts, reusable templates that help structure how the model uses the server.
In practice, the assistant asks the server for its list of tools (tools/list), the model picks one that fits the person's request, and the assistant runs it (tools/call). Each tool comes with a name, a description and a schema that says what inputs it needs, so the model can work out when and how to use it.
Servers can run on the person's own computer or on the internet. A server on the internet is called a remote server and uses a transport called Streamable HTTP. That is the kind a website or store would offer.
Who governs MCP now?
The Agentic AI Foundation (AAIF), part of the Linux Foundation, hosts MCP.
The Linux Foundation announced the AAIF on 9 December 2025, with MCP contributed by Anthropic, goose contributed by Block, and AGENTS.md contributed by OpenAI. Its platinum members at launch were Amazon Web Services, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI.
In the same announcement, the Linux Foundation said there were more than 10,000 published MCP servers and that MCP had been adopted by Claude, Cursor, Microsoft Copilot, Gemini, VS Code and ChatGPT. Those figures are the Linux Foundation's, and we have not checked them independently.
The specification has its own governance too. The November 2025 revision formalised MCP's governance structure, working groups and interest groups.
What changed in the MCP specification in 2026?
The 28 July 2026 revision made MCP stateless, which makes remote servers simpler to run. These are the main changes listed in the official changelog:
- No more sessions or handshake. The revision removes protocol-level sessions and the
initializehandshake. Every request now carries its protocol version and the client's capabilities. - A new discovery call. Servers must implement
server/discover, which advertises their supported versions, capabilities and identity. - Caching hints. Results from list calls such as
tools/listmust carry a freshness hint and a cache scope, so clients can cache them instead of asking again. - Deprecated features. Roots, Sampling and Logging are deprecated. They still work during a deprecation window of at least twelve months, but new servers should not add them.
- Tasks moved to an extension. Long-running tasks, added experimentally in November 2025, became an official extension.
The previous revision, dated 25 November 2025, added URL mode elicitation, experimental tasks and changes to authorization.
If you run an MCP server, check which revisions your clients support. Claude's documentation, for example, lists the authorization specifications it follows, so client support can lag behind the newest spec.
What does MCP mean for a website or online store?
It is one of the ways an AI assistant can use your business directly, instead of only reading your pages.
An assistant reading your site has to guess from the page text. An assistant connected to your MCP server can call a tool like "search products" or "check order status" and get a structured answer. Commerce standards build on this. UCP, the checkout standard behind Google's AI surfaces, lists MCP as one of its transports, alongside REST, A2A and an embedded option. See what UCP is.
How people reach a server today:
- In Claude, a user or organisation owner can add a remote server as a custom connector by entering its URL, or install one from Anthropic's reviewed directory.
- In ChatGPT, OpenAI's documentation describes adding a custom MCP server through plugins.
Security deserves attention. OpenAI's MCP documentation says custom servers are not developed or verified by OpenAI, warns that data a server can access could carry prompt injections (hidden instructions), and recommends connecting only to servers whose underlying application you know and trust. If you run a server, keep write actions such as orders behind confirmation and clear limits.
For a practical walkthrough of whether your store needs a server, and what Shopify and WooCommerce already provide, read does your store need an MCP server?. For where MCP fits among the other standards, see agentic commerce protocols explained.
How Sill helps
Sill publishes the files AI agents read, including an MCP server for your site, and keeps them current. It shows you which AI agents visit and what they ask for. When an agent places an order, your rules approve it, decline it or hold it for a person to review, and every decision is signed and recorded. Run a free readiness scan to see what agents find on your site today.
Questions people ask
Is MCP only for Claude?
No. Anthropic created it, but the MCP documentation lists support in ChatGPT, Visual Studio Code, Cursor and other tools, and the Linux Foundation says Microsoft Copilot and Gemini have adopted it too.
Is MCP a payment protocol?
No. MCP describes how an assistant calls tools. A tool can start a purchase, but payment proof and checkout rules come from other standards such as UCP, ACP or AP2. UCP lists MCP as one of its transports, and the AP2 site describes AP2 as an extension for A2A and MCP.
Is it safe to connect an AI assistant to any MCP server?
Not automatically. OpenAI's documentation says custom MCP servers are not developed or verified by OpenAI, warns that data a server can access could contain prompt injections (hidden instructions), and recommends connecting only to servers whose underlying application you know and trust.
What is the difference between MCP and an API?
An MCP server usually sits on top of an existing API. The difference is that MCP describes tools in a standard way an AI model can read and choose from, so one assistant can use many services without custom code for each.
Sources
- Anthropic: Introducing the Model Context Protocol · checked 16 Nov 2026
- Model Context Protocol: What is MCP? · checked 16 Nov 2026
- Model Context Protocol: Architecture overview · checked 16 Nov 2026
- MCP specification 2025-11-25: Key changes · checked 16 Nov 2026
- MCP specification 2026-07-28: Key changes · checked 16 Nov 2026
- Linux Foundation: Formation of the Agentic AI Foundation · checked 16 Nov 2026
- OpenAI developers: MCP · checked 16 Nov 2026
- Claude docs: Build an MCP server for Claude · checked 16 Nov 2026
- UCP specification overview · checked 16 Nov 2026
- AP2: Agent Payments Protocol · checked 16 Nov 2026