MCP Server Documentation: The Complete Index

MCP server documentation is scattered across half a dozen domains. The official spec lives in one place, the reference implementations in another, the language SDKs in a third, and each AI host has its own config guide. This page is the index: what to read, where it lives, and what you actually need from each source.

The short answer

Five canonical sources cover everything you need:

1. The spec: modelcontextprotocol.io. Wire format, primitives, handshake, transports.
2. The reference servers: github.com/modelcontextprotocol/servers. MIT-licensed, Anthropic-maintained, production-ready templates.
3. The language SDKs: Python and TypeScript, both documented at modelcontextprotocol.io with full API references.
4. Host config guides: Claude Code, Cursor, Windsurf, and OpenCode each ship their own MCP setup docs. Same JSON shape, different file paths.
5. Vendor MCP server docs: AWS, Azure, Google Cloud, Cloudflare, Stripe, Figma, Zerodha. Each lives on the vendor's own developer portal.

For the protocol overview before reading any docs, see What is Model Context Protocol. For the server side specifically, see What is an MCP server. The rest of this page tells you what each source actually contains, who should read it, and the quickest path through.

The official spec: modelcontextprotocol.io

The canonical specification. Maintained by Anthropic with community contributions. This is where you go to answer questions like:

• What does a tool definition look like in JSON?
• How does the handshake between host and server work?
• What error codes does the spec define?
• What does the resources primitive support and what does it not?
• What is the difference between stdio and HTTP+SSE transports?

Read the "Concepts" section first if you are new to MCP. It explains tools, resources, and prompts in roughly 5 minutes each. Then read the "Specification" section if you are implementing your own server or client. The spec is concise; the whole document is shorter than a typical RFC.

What the spec does not cover: best practices, performance tuning, real-world failure modes, security advisories. For those, the reference implementations and host docs are more useful.

Reference implementations: github.com/modelcontextprotocol

The single most useful repository for MCP work. Anthropic publishes MIT-licensed reference servers covering the most common needs:

• filesystem - read and write local files with scoped permissions
• git - inspect git history, blame, status, diffs
• github and gitlab - repo metadata, PR review, issue management
• postgres and sqlite - schema introspection and read-only queries
• fetch - HTTP GET to a URL with simple HTML-to-text extraction
• memory - knowledge-graph-style notes the model can write to and read back
• slack, brave-search, and a long tail of integrations

Each server has its own README with the npx install command, the required environment variables, and the list of tools it exposes. Read these even if you are not installing them: they are the stylistic template for how to write good MCP server docs.

The repository also has the official Python and TypeScript SDKs (under /python-sdk and /typescript-sdk) plus example projects that show how to author a new server from scratch.

Language SDK documentation

Two first-party SDKs covering the languages most MCP authors reach for first.

Python SDK

Documented at modelcontextprotocol.io under the Python guide. Covers the FastMCP server framework (decorator-based, minimal boilerplate), low-level Server class for full control, async tool handlers, resource templates, and prompt definitions. Best for data-science MCP servers, AI orchestration tools, and anything that wraps existing Python libraries.

TypeScript SDK

Documented under the TypeScript guide. Class-based API, first-class type definitions for tool input schemas via Zod, built-in stdio and HTTP transports. Best for npm-installable MCP servers (most of the ecosystem ships as npm packages and uses npx -y for distribution).

Community SDKs

Rust, Go, C#, and Kotlin SDKs exist as community projects. Listed at modelcontextprotocol.io/community-sdks with maturity flags. Pick a first-party SDK unless you have a specific runtime constraint that rules out Python and TypeScript.

Host-specific config guides

Each MCP-aware host ships its own config docs. The protocol is identical across all of them; what differs is where the host reads its config from and what CLI helpers it provides.

Claude Code

• Docs: code.claude.com/docs/en/mcp
• Config files: .mcp.json (project), ~/.claude.json (user), local overlay
• CLI helpers: claude mcp add, claude mcp list, claude mcp remove

For a deeper walkthrough including the three config scopes and common errors, see Claude Code MCP.

Cursor

• Docs: cursor.com/docs/mcp
• Config files: .cursor/mcp.json (project), ~/.cursor/mcp.json (user)
• CLI helpers: edit JSON directly, no add command shipped at time of writing

Windsurf (Codeium)

• Docs: docs.codeium.com/windsurf/mcp
• Config file: ~/.codeium/windsurf/mcp_config.json (single global)
• CLI helpers: edit JSON directly

OpenCode

• Docs: opencode.ai/docs/mcp-servers/
• Config file: opencode.json in project, with mcp object using the same JSON shape as Claude Code and Cursor
• Notes: OpenCode is open source (GitHub: opencode-ai/opencode), MIT-licensed, and supports 75+ models. As of 2026 they have around 6.5 million monthly developers and full MCP support is a first-class feature, not a plugin.

Other hosts

• Continue.dev - docs.continue.dev/customize/mcp
• Zed - zed.dev/docs/assistant/model-context-protocol
• Cline - GitHub README under cline/cline plus the VS Code extension docs
• Visual Studio Code - code.visualstudio.com/docs/mcp (Microsoft now owns three SERP slots for "mcp server")

Vendor MCP server documentation

Most major cloud and SaaS vendors now publish official MCP servers alongside their existing API documentation. As of May 2026:

• AWS - aws.amazon.com/mcp. General availability hit in May 2026; the AWS MCP server is now in the Bing top 10 SERP for "mcp server".
• Microsoft Azure - learn.microsoft.com/azure/mcp. Plus Visual Studio Code MCP integration at code.visualstudio.com/docs/mcp.
• Google Cloud - cloud.google.com/mcp. Plus Stitch (their AI design MCP) at stitch.google.com. Both shipped April 2026.
• Cloudflare - developers.cloudflare.com/mcp. Workers, KV, R2, and DNS exposed as MCP tools.
• Stripe - docs.stripe.com/mcp. Read-only by default; payments require explicit scope.
• Figma - help.figma.com/mcp. Dev Mode MCP for design-to-code workflows; ~4 million Figma users now have an MCP entry point.
• Zerodha - kite.trade/docs/mcp. Kite Connect MCP server for India's largest broker (30 million retail trading clients). India is now the number-one Bing market for "what is mcp server".
• Anthropic itself - docs.anthropic.com/mcp. Cross-product MCP guidance covering Claude Code, Claude Desktop, and the Anthropic API.

AgentDrop docs as a worked example

AgentDrop publishes its own MCP server documentation at docs.agent-drop.com/guides/mcp-server. It is included here because it is one of the few MCP servers covering agent-to-agent communication specifically (every other server in the ecosystem connects an AI model to a service; AgentDrop connects two AI agents to each other).

What you find there:

• One-line install for Claude Code, Cursor, Windsurf, and OpenCode
• Tool reference for send_file, check_inbox, download_transfer, list_connections, and the rest
• End-to-end walkthroughs: a Claude Code instance sending a file to a Cursor instance on a different machine; a Python script using the SDK directly; an MCP host triggering a transfer to a human via email
• The cryptography reference (X25519 ECDH, AES-256-GCM, HKDF-SHA256) with a separate page at encrypted file transfer for AI agents that walks the maths

The point of including this here is not promotional. It is a template you can copy. The structure (install / tools / walkthroughs / security model) is what good MCP server docs look like, regardless of who publishes them.

How to read MCP docs effectively

MCP documentation is split across many domains because the protocol itself is intentionally vendor-neutral. There is no single all-knowing source. Instead, every audience reads a different subset.

If you are a Claude Code, Cursor, or Windsurf user

Start at your host's config docs. You almost never need the spec. Read the host-specific guide, install three servers from the directory, and you will understand 90% of what you need.

If you are authoring an MCP server

Read the spec's "Concepts" section, then pick a language SDK. The reference implementations in github.com/modelcontextprotocol/servers are the best template. Start by copying the closest server to what you want to build, then trim and modify.

If you are integrating MCP into a host

Read the full spec including the "Specification" section. Read at least one TypeScript SDK source file to see how the message handling looks in practice. Start with stdio transport; HTTP+SSE adds complexity you can defer.

If you are evaluating MCP for a buying decision

The spec's "Concepts" section plus this page's vendor list is enough. The protocol is now multi-vendor and enterprise-mainstream; the choice is no longer whether to adopt MCP but which servers to install first.

FAQ

Where do I find official MCP documentation?

The canonical specification lives at modelcontextprotocol.io, maintained by Anthropic and the MCP community. The spec covers the wire format (JSON-RPC 2.0), the three primitives (tools, resources, prompts), the handshake, and the two transports (stdio and HTTP+SSE). Reference servers and language SDKs live at github.com/modelcontextprotocol.

Where are the reference MCP servers documented?

github.com/modelcontextprotocol/servers. Anthropic maintains MIT-licensed reference implementations for filesystem, git, GitHub, GitLab, Postgres, SQLite, fetch, memory, Slack, and more. Each server has its own README with install commands, environment variables, and usage examples.

Which MCP SDKs have official documentation?

Two first-party SDKs: Python and TypeScript, both with full API references at modelcontextprotocol.io. Community SDKs in Rust, Go, C#, and Kotlin are listed but vary in maturity. Pick first-party unless you have a specific runtime constraint.

Where are host-specific config guides for Claude Code, Cursor, and Windsurf?

Claude Code: code.claude.com/docs/en/mcp. Cursor: cursor.com/docs/mcp. Windsurf: docs.codeium.com/windsurf/mcp. OpenCode: opencode.ai/docs/mcp-servers/. The protocol is identical across hosts; only file paths differ.

Are vendor MCP servers from AWS, Microsoft, and Google documented?

Yes. AWS MCP at aws.amazon.com/mcp (GA May 2026, now in Bing top 10 SERP). Microsoft Azure at learn.microsoft.com/azure/mcp plus Visual Studio MCP integration at code.visualstudio.com/docs/mcp. Google Cloud at cloud.google.com/mcp plus Stitch at stitch.google.com. Cloudflare, Stripe, Figma, and Zerodha all ship their own MCP docs at their developer portals.

How do I document my own MCP server?

Three things people need: the install command (one line, copy-paste ready), the required environment variables (with example values), and the list of tools your server exposes (each with a one-sentence description). The Anthropic reference servers are the best stylistic template. AgentDrop publishes its own MCP docs at docs.agent-drop.com/guides/mcp-server as another worked example.

Where to go next

Related guides on agent-drop.com:

• What is Model Context Protocol - the protocol overview every doc on this page assumes
• What is an MCP server - the server-side mechanics if you are about to install your first one
• Claude Code MCP - deeper walkthrough of the Claude Code config layer specifically
• MCP servers directory - curated list grouped by what they do, with install commands
• MCP server for file transfer - worked example of one specific MCP server end-to-end

To run AgentDrop as an MCP server in any of the hosts above, the quickstart gets you running in about 60 seconds. Free tier: 50 transfers and 50 MB files per month, no card required. The MCP server itself is MIT-licensed and the source lives at github.com/qFlav/agentdrop-mcp-server.