Tool use

Tool use (also called function calling) is the mechanism that turns a text model into a doer.

The idea: the model can’t actually compute 2^100, check the weather, or read your disk — but it can emit a structured request like:

{ "tool": "calculator", "args": { "expr": "2 ** 100" } }

The surrounding software (the harness) runs the real thing, hands the result back to the model, and the model weaves it into the answer. From the model’s point of view, tools are just another kind of text in and out — that’s what makes them so flexible.

Where you’ve already met this:

• Search-grounded answers — the “search” tool.
• Code agents — the shell tool, the file tools.
• MCP (Model Context Protocol) — an open standard, think “USB-C for tools”, so one set of tool servers can plug into many AI apps.