Function Calling#
Function Calling lets a model request an action from your application. MEGA returns the requested tool call; your application validates it, decides whether it is allowed, performs the action, and sends the result back in the next model request.
Use Function Calling only with a route whose
supports_tools capability is enabled in the live
Model Catalog.
Define a narrow tool#
Describe the smallest action that the model needs. Require the fields you need, disallow unexpected fields where possible, and avoid tools that accept a raw shell command or unrestricted URL.
{ "type": "function", "function": { "name": "get_weather", "description": "Read the current weather for a city", "parameters": { "type": "object", "properties": { "city": {"type": "string"} }, "required": ["city"], "additionalProperties": false } }}Include this object in the standard OpenAI-compatible tools array for
Chat Completions or use the
equivalent tool field for Responses.
Handle a tool call safely#
- Read the tool name and arguments from the model response.
- Validate the name, argument types, and business rules in your application.
- Check that the signed-in user is authorized for the requested action.
- Execute the action with scoped credentials.
- Return a concise tool result in the next message or Responses input.
Treat model output as untrusted input. A valid schema does not prove that an action is appropriate, safe, or authorized.
Design for retries#
Use idempotency keys or an application-side confirmation step for actions that send messages, make purchases, modify data, or trigger Jobs. A user may retry after a network failure, and an application must avoid performing the same side effect twice.
For a complete Chat request, see Chat Completions. For agent-style input and events, see Responses API.