Model Context Protocol (MCP)
The Model Context Protocol (MCP) is an open standard that lets AI assistants connect to external tools and data sources through one consistent interface instead of a bespoke integration per pairing.
MCP standardises three things: how a tool describes itself, how it is called, and how results and errors are returned. An MCP server exposes capabilities; an MCP client — the assistant — reads the available tools and decides when to call them. Announced by Anthropic in late 2024, the protocol was handed to the Linux Foundation in late 2025.
Because tool descriptions are read by the model, they behave as part of the prompt: a vague description produces a tool called at the wrong moments. Anything an MCP server returns must be treated as untrusted data rather than as instructions the assistant may follow.
How it works
An MCP server wraps something useful, such as a database, a file system or a company's API, and advertises the tools and resources it offers. An MCP client inside an assistant connects to one or more servers, lists what they offer and passes those descriptions to the model. When the model decides to use a tool, the client sends the call to the right server and returns the result into the conversation.
Because the interface is standard, a server written once works with any assistant that speaks MCP, and an assistant gains new abilities by connecting another server rather than by being rebuilt.
Example
An assistant connected to a Search Console MCP server can pull a site's queries and index status on request; connected to a file-system server as well, it can read the site's code and propose the fix for a page that is not being indexed. Neither vendor had to build an integration with the other.
Common mistakes
- Installing community MCP servers without reading what they can access.
- Writing vague tool descriptions, which make the model call tools at the wrong moments.
- Treating text a server returns as trusted instructions rather than as data.
