MCP (Model Context Protocol)

Anthropic open-sourced the Model Context Protocol on November 25, 2024, as an open standard for connecting AI assistants to the systems where data lives, replacing a custom integration per data source with one protocol. Developers expose data and tools through MCP servers, and AI applications connect to them as clients. The launch included the specification and SDKs, local server support in the Claude desktop apps, and pre-built servers for Google Drive, Slack, GitHub, Git, Postgres and Puppeteer (Anthropic, Introducing the Model Context Protocol).

Under the hood, the specification uses JSON-RPC 2.0 messages between hosts (the LLM applications), the clients inside them, and servers. Servers can offer resources, prompts and tools; clients can offer servers sampling, roots and elicitation. The design takes inspiration from the Language Server Protocol, which did the same job for programming-language support in editors (MCP specification). Support now extends beyond Anthropic; OpenAI's Agents SDK, for example, can connect agents to MCP servers (OpenAI Agents SDK, Model context protocol).

The security model matters as much as the interface. The specification says tools represent arbitrary code execution, that tool descriptions should be treated as untrusted unless they come from a trusted server, and that hosts must get explicit user consent before invoking a tool. OpenAI's documentation gives the same advice from the client side: connect only to servers you trust, because they can see data from the model's context and act with the credentials you provide. What can MCP do? and MCP vs custom tool integration cover the practical choices, and excessive agency covers limiting what connected tools can do.