Tool calling
Send tool declarations and handle returned tool calls with an OpenAI-compatible model request.
Tool calling
Tool calling lets the model request an application-owned action. Your application remains responsible for validating arguments, authorizing the action, running it, and returning the result.
Never let a model-generated tool call bypass authentication, spend limits, filesystem boundaries, or human approval. Treat the returned arguments as untrusted input and validate them against the tool's schema before execution.
Use the selected model's current capability metadata and completion response to verify tool support.
Next: Chat and Work.