MCP (Model Context Protocol)¶

Intermediate → Advanced · 7 topics

The AI Agents series showed MCP in one section. This series goes all the way: the protocol itself, building and testing servers, connecting them to real hosts and frameworks, running them remotely with proper auth, and the security problems that come with letting other people's tools into your model's context. MCP shows up in more and more GenAI job descriptions, and "have you built an MCP server?" is now a common interview question.

# Pages 1, 2 and 7 run offline with the Python standard library.
# Pages 3-6 need packages and are marked "no-run"; their outputs are from real runs:
pip install "mcp>=2"                                  # the official Python SDK (servers and clients)
pip install langchain-mcp-adapters openai-agents      # page 6, in a separate environment (they still use mcp 1.x)

#	Topic	Sub-topics
1	What MCP is	The N×M problem · host, client, server · tools, resources, prompts · JSON-RPC · handshake era vs the stateless 2026-07-28 spec · MCP vs function calling
2	The protocol from scratch	A stdlib server and client · `server/discover` · tools and resources · protocol errors vs `isError` · multi round-trip elicitation · the stdio transport
3	Building servers with the Python SDK	`MCPServer` · structured output · `ToolError` · resources and templates · prompts · progress · `Resolve` and `Elicit` · lifespan · MCP Inspector
4	Clients and hosts	Claude Code, Claude Desktop, Cursor, VS Code configs · the Python `Client` · MCP tools as LLM tools · an agent loop over MCP · several servers · caching
5	Remote servers and auth	Streamable HTTP on the wire · `Mcp-Method` / `Mcp-Name` headers · stateless scaling · mounting in FastAPI · OAuth 2.1 · no token passthrough
6	MCP in agent frameworks	`langchain-mcp-adapters` · LangGraph · OpenAI Agents SDK · the Claude API MCP connector · Claude Code · choosing an approach
7	Security and production	Tool poisoning · rug pulls and pinning · injection through results · the lethal trifecta · approvals · audit logs · testing · checklist

flowchart LR
    U[User] --> H[Host<br/>Claude, IDE, your agent]
    H --> L[LLM]
    H --> C1[MCP client] <-->|stdio| S1[Local server<br/>files, git, CLI]
    H --> C2[MCP client] <-->|Streamable HTTP + OAuth| S2[Remote server<br/>company API]
    S1 --> T1[tools · resources · prompts]
    S2 --> T2[tools · resources · prompts]

A fast-moving standard

These pages follow the 2026-07-28 MCP spec and the Python SDK mcp 2.x (where FastMCP became MCPServer). Older tutorials use the initialize handshake and FastMCP; page 1 explains both, and the SDKs still talk to older servers and clients.

Before you start

These pages assume structured output & tool calling, ReAct from scratch and Tools and MCP. Some async Python helps for pages 3-6: see async programming.

Back to: Notes overview