Mastering Model Context Protocol (MCP)

You can write a Python function and you have used an AI assistant. You have never built a tool for one. This book takes you from that line to a server that a real AI app calls correctly, tests, ships and runs behind a gateway for a whole team.

It teaches the Model Context Protocol against the current specification, dated 28 July 2026. MCP was a session protocol: a client opened a connection, shook hands once, and held state. The July 2026 revision is stateless. Requests are self-contained, capabilities travel with every request, servers no longer initiate calls. Many tutorials written before July 2026 still teach the old shape, and the classic first import no longer works. This book teaches the new one, explains every line, and shows you how to tell which revision any server speaks.

One running case: an on-call engineer at a fictional parcel company, one night of checkout logs, and a server called opslog that grows one chapter at a time from a single tool to four typed tools, two resources and a prompt, with tests, host configuration, security, a gateway and a container. The data script is printed in full, so every number you check comes out the same.

Fifteen chapters, eighteen diagrams, graded exercises with check values and worked answers, a glossary, a quick reference and a troubleshooting index:

• Why hand-written tool wrappers break, and what the protocol absorbs


• Host, client and server, and which one you are actually writing


• JSON-RPC, HTTP and the stateless July 2026 revision, read off the wire


• The two transports, and why nothing else may go to stdout


• A reproducible project with uv, and a server that runs in a dozen lines


• How type hints and docstrings become a schema a model calls correctly


• Resources and prompts, and the cost of making everything a tool


• Tests, the Inspector, and reading the log when a call fails


• Plugging the server into Claude Code, Claude Desktop, VS Code and Cursor


• Your own client and a LangChain agent on a scripted stand-in model, with no API key needed


• Community and hosted servers, vetted and pinned


• Streamable HTTP with a token, and the pre-exposure checklist


• The gateway pattern: many servers behind one URL


• A wheel, a dry-run publish and a container with everything inside


• A capstone: a second server, designed and shipped from a one-line brief

Python 3.12 and fastmcp 4.1.0. Every listing, and the output printed under it, was run on the versions the front matter names, except where a chapter says it was not: the Claude Desktop, VS Code and Cursor files come from their makers' documentation, and install commands are quoted from the tools' own pages.