Reference

MCP documentation

The official MCP documentation lives at modelcontextprotocol.io. It covers the protocol specification, SDK references, host implementation guides, and the three primitives. Versioned reference (current at writing: 2025-06-18), maintained under LF AAIF governance.

Documentation surfaces

Protocol spec

At modelcontextprotocol.io/specification. The authoritative reference for client and server implementers building from scratch.

SDK reference

Python, TypeScript, and additional language SDKs handle the JSON-RPC wire mechanics and capability negotiation for you.

Quickstarts

Getting-started guides for first-time server and client builders. Run a reference server in Claude Desktop or another host inside an hour.

Architecture overview

Host, Client, Server actors. Capability negotiation. The three primitives. Where to dig in based on which side you're building.

Reading order for new builders

  1. Run a quickstart. Install a reference server (filesystem, GitHub, or similar) in a host like Claude Desktop. See the wire in action before reading the spec.
  2. Read the architecture overview. Internalize the three primitives (Resources, Tools, Prompts) and the three client capabilities (Roots, Sampling, Elicitation).
  3. Read the spec sections for your role. Server builders focus on Resources, Tools, Prompts. Client builders focus on Roots, Sampling, Elicitation plus the initialize handshake.
  4. Read the SDK reference for your language. The SDK abstracts JSON-RPC wire mechanics; you write your tool functions in idiomatic Python, TypeScript, or your language of choice.
  5. Read the security guidance. Tool Poisoning, Cross-Server Shadowing, Rug Pulls, and Return Value Injection are spec-level attack patterns every implementer should understand.

Gaps in the official docs

The mitigation guidance for the four spec-level attack patterns is still maturing in the official docs. Best-practices sections are being added as the threat landscape develops. For now, the AgentSeal 2025 server scan (66% of 1,808 surveyed servers had findings) and the Invariant Labs WhatsApp tool-poisoning demonstration give the operational picture.

Community resources

Beyond the official docs: GitHub awesome-mcp lists, vendor deep-dives from Anthropic, Cloudflare, Stripe, and others, plus Hacker News threads from CVE disclosures and architectural discussions. The community resources catch operational nuance the spec doesn't formalize.

Related on MCPowered