Tool definitions differ across OpenAI, Anthropic, Gemini: get.weather only works in Gemini

The guide lays out how tool calling works: the model returns a request to run a function with specific arguments, and your own code executes it. What changes between vendors is only the packaging of that request — where the name, description and argument schema sit, and which characters a name may contain. OpenAI's two endpoints already differ from each other, with Chat Completions nesting fields under function and Responses placing them at the same level, while Anthropic renames the schema field to input_schema and Gemini gathers declarations under functionDeclarations.

The guide also flags pre-send traps. OpenAI's strict mode requires additionalProperties set to false and every property listed as required, and the default for strict differs between Chat Completions and Responses. Anthropic rejects the whole request with a 400 error if even one input_examples sample does not match the tool's own schema, and those samples add prompt tokens — roughly 20–50 for simple ones and 100–200 for deeply nested shapes. Gemini forbids specifying parameters and parametersJsonSchema at the same time.

On the local side, Ollama labels each model with capabilities such as tools, thinking, vision and embedding, and only models bearing the tools label can call functions at all. Because the library changes often, the guide advises checking the label on the tag page you plan to use rather than assuming a model size implies support.