Tools and schemas

A LangChain tool pairs executable code with a schema. The model sees the tool’s name, description, and arguments; your application executes the function.

Tiny example

from langchain.tools import tool

@tool
def get_order_status(order_id: str) -> str:
    """Return the shipping status for one order the caller may access."""
    return order_service.lookup(order_id)

Type hints define the input schema. The docstring helps the model decide when the tool applies.

A useful tool contract

1. Narrow name: get_order_status, not do_everything.
2. Clear description: say when to use it and what it returns.
3. Typed arguments: reject missing or malformed input early.
4. Application checks: authenticate the caller and authorize the specific record.
5. Small output: return what the next step needs, not an entire database row.

Failure note

Schema validation is not authorization. The model can request order_id="someone-elses-order"; the tool must enforce tenant and object permissions. Do not expose broad filesystem, SQL, or shell access because the description asks the model to be careful.

Related

Tool calling · create_agent · Middleware