Model Context Protocol Documentation

What MCP is

MCP defines how an AI application (the host, running one or more clients) talks to servers that expose capabilities. Before a standard existed, every integration was custom to one assistant. With MCP, a server you write once can be used by any compatible client. The documentation compares it to a USB-C port for AI applications.

Messages use JSON-RPC, carried either over standard input and output for local servers or over HTTP for remote ones.

The concepts to get right first

A server can offer three kinds of capability:

• Tools: actions the model chooses to call, such as creating a ticket or querying a database.
• Resources: data the application or user attaches as context, such as a file or a record.
• Prompts: reusable templates a user can pick.

Most design mistakes come from making everything a tool. Deciding what the model should invoke and what the user should attach is most of designing a good server.

A reading order that works

1. The introduction and the architecture overview.
2. Understanding servers and understanding clients.
3. The build-a-server tutorial in your language, then connect it to a local client to see it working.
4. The MCP Inspector guide, which lets you call your server's tools and view responses without an AI client in the loop. Use it constantly while developing.
5. Authorisation (OAuth 2.1 for remote servers) and the security best practices page before exposing anything beyond your own machine.

The specification is versioned by date, so check which version your client supports.

SDKs and choosing one

As of September 2026, the reference servers repository on GitHub had about 90,500 stars and the specification repository about 9,300, a sign of how widely the protocol has been adopted.

Official SDKs exist for TypeScript, Python, C#, Go and Rust (the most complete tier), with Java, Ruby, Swift, PHP and Kotlin also available. The protocol is small enough that you could implement a simple server directly, but an SDK saves you from edge cases in transport and lifecycle handling.

Tips for a first server

• Start with two or three well-named tools around one real task, not a wrapper for every endpoint of an API.
• Write tool descriptions as carefully as a prompt; the model relies on them.
• Return concise, human-readable results rather than raw payloads.
• Read Anthropic's Writing Effective Tools for AI Agents alongside the docs.

Common questions

Is MCP only for Claude? No. The documentation lists support in ChatGPT, VS Code, Cursor and many other clients.

Do I need to build a server, or can I use existing ones? Many ready-made servers exist for common tools. Build your own when you need access to your own systems or data.

Is it secure? The protocol supports OAuth-based authorisation for remote servers, but security depends on your implementation. Read the security best practices page, and be careful which servers you install, since a server can run code on your machine.

How long does a first server take? Following the tutorial with an SDK, a simple working server is an afternoon's work.

Best for: engineers connecting AI assistants to internal systems, and anyone building developer tools for AI clients.