Defining tools with MCP

Why it matters

An MCP server is worth its operating cost only if its tools are as clear to Claude as a direct tool definition. 3.7 asks when that cost is justified; this lesson shows what the server side of the choice involves.

What to practise

Create a FastMCP server over an in-memory dictionary of documents. Define read_doc_contents and edit_document with the @mcp.tool decorator, giving each a name and description and describing every argument with Pydantic’s Field. Note that the SDK generates the schema from the type hints, and that a missing document raises a ValueError. The edit tool’s exact-match old_str shows how an argument description constrains Claude’s input.

How it fits

The decorator produces the same name, description and input_schema that 3.7.1 compares across integration mechanisms. Weigh it with 3.7.2: if only one application calls these tools, a direct custom tool carries the same definition without a server to run.