The structured definition of a function (name, description, JSON-schema parameters) that an LLM can call — the contract that lets the model decide when and how to invoke external capabilities.
A Tool Schema describes a callable function in a format the LLM can reason about: a name, a natural-language description (which the model reads to decide when to use it), and a JSON Schema for parameters. OpenAI, Anthropic, and Google use compatible-but-not-identical formats. The description is the most important part — a vague description means the model misuses or ignores the tool. Pydantic is the de facto Python library for generating tool schemas from typed function signatures.
Defining a `create_invoice(customer: str, amount: float, due_date: date)` tool schema so an LLM customer-service agent can issue invoices when asked.
Tool Schema quality directly determines agent quality — well-written descriptions and minimal parameter sets get used correctly; sprawling or ambiguous schemas confuse the model.
Need help implementing this in your business?
Get Started