What is the Model Context Protocol (MCP)?

The Model Context Protocol (MCP) is an open standard for connecting AI applications to outside data, tools and workflows through one common interface.

How it works

The official docs compare MCP to "a USB-C port for AI applications": instead of every assistant building a custom connector for every service, both sides speak one protocol (MCP docs).

There are two roles:

• MCP servers expose capabilities. A server might offer tools (actions such as "create ticket" or "search orders"), resources (data such as documents or records) and prompts (reusable instructions).
• MCP clients live inside AI applications. The docs list assistants such as Claude and ChatGPT and developer tools such as Visual Studio Code and Cursor among the clients that support MCP. The client discovers what a server offers and lets the model use it during a conversation.

The protocol is versioned by date. As of October 2026 the current version is 2026-07-28; a new date marks the last backwards-incompatible change (MCP versioning).

MCP started at Anthropic. On December 9, 2025, Anthropic donated it to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI. The post announcing the move says the maintainers keep control of the protocol's technical direction (MCP blog).

Why it matters for your website

MCP matters if people want to use your product or your knowledge from inside their own AI assistant. A software company can publish an MCP server so a customer's assistant can search its documentation or read account data with the customer's permission, instead of copying and pasting from your site.

MCP is not the same as WebMCP. An MCP server runs outside the browser and a user or admin connects it to their assistant. WebMCP tools live inside a web page and are used by an agent in the visitor's browser tab.

Every tool you expose acts with real permissions, so design it like an API: give it the narrowest scope that works, prefer read-only tools, and plan for prompt injection, where text the model reads tries to make it misuse a tool.

Example

A SaaS vendor publishes a read-only MCP server with two tools: search_docs and get_plan_limits. A customer's admin connects it to the company's AI assistant. When an employee asks, "Does our plan include single sign-on?", the assistant calls get_plan_limits, reads the documented answer and replies with a link to the docs page, rather than guessing from what the model happened to learn in training.

Related terms

• WebMCP: a proposed browser version of the same tool idea, for web pages.
• AI agent: the kind of software that calls MCP tools.
• Large language model (LLM): the model behind the AI application.
• Prompt injection: the main reason to keep tools narrow.