Tool Calling

Tool calling uses OpenAI-compatible or native protocol fields supported by the current model. First confirm tool capability in GET /v1/models/{model} and Model Market, then send tools and a tool selection strategy in the request.

Tool arguments from a model are untrusted input. Validate arguments, execution permissions, and call results before returning the result to the model in a later message. Model capabilities and field compatibility can change; use current response fields, Model Market, and actual tests as the source of truth.