a hands-on course

Build an
MCP Server

(Model Context Protocol)

Package your own tools and data as a server any AI app can plug into — in Python, in about half an hour.

A free, hands-on tutorial: build a Model Context Protocol (MCP) server in Python. Use the official SDK's FastMCP to expose tools, resources and prompts — your type hints become the schema — then connect it to Claude Code and the Claude API so a real AI can call it. The sequel to Build an AI Agent: there you gave one agent your tools; here you package them so any app can. Building the server needs no API key.

in the real world

MCP is becoming the standard way AI apps connect to the real world — the same server works in Claude Code, Claude Desktop, the API and beyond. Build one and your tools and data are available to every AI tool that speaks the protocol, not just one script.

Start with why MCP exists →

Part 1 — Why MCP, and your first server

Why MCP exists

Every AI app needs your tools and data, and each one wired them up differently. MCP is the one plug that fits them all.

A server in 15 lines

Install the SDK, write one tool with FastMCP, run it, and poke at it live in the MCP Inspector — no protocol code.

Part 2 — The three primitives

Tools: let it do things

Tools are the actions your server offers. Add a few real ones and see how type hints and docstrings become the interface.

Resources & prompts

The other two primitives: resources hand data to the model (GET-like), prompts are reusable templates users can invoke.

Part 3 — Connect it & ship

Connect it to Claude

Point Claude Code at your server with one command — then the Claude API — and watch it call your tools for real.

Secure it & ship

A server exposes your data, so lock it down — plus the v2 that's coming, how to pin it, and where to go next.