Building an MCP server is genuinely small — you write plain Python functions and let the framework expose them as tools. Here's the shortest useful version using fastmcp.

1. Install

pip install fastmcp

2. Define a tool

Each tool is a function; its signature and docstring become the schema the model reads:

from fastmcp import FastMCP

mcp = FastMCP("Weather")

@mcp.tool
def get_forecast(city: str) -> str:
    """Return a short weather forecast for a city."""
    # ... call your real weather API here ...
    return f"Sunny in {city}, 21C."

if __name__ == "__main__":
    mcp.run()

3. Run and connect it

Run python server.py and point an MCP client at it (locally over stdio, or expose it over HTTP for remote clients). The client lists get_forecast and the model calls it when a question needs a forecast.

Good habits

Write clear tool names + docstrings — that's the model's only guide to when to use them.
Mark or separate anything that writes/changes data; keep read-only tools obviously safe.
Validate inputs and return concise, text-friendly results (models read them as context).

That's the whole shape of it. The other half is the thing that calls your server — see How to build an MCP client in Python, or What is an MCP server? for the concepts.