Tool calling

Function/tool calling passes through the gateway untouched — define tools in the format of the surface you call, and the model's tool_calls come back in the same format, whichever provider actually served the request.

Define tools (OpenAI surface)

{
  "model": "claude-opus-4-7",
  "messages": [{"role": "user", "content": "What's the weather in Osaka?"}],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_weather",
        "description": "Get current weather for a city",
        "parameters": {
          "type": "object",
          "properties": {
            "city": {"type": "string"}
          },
          "required": ["city"]
        }
      }
    }
  ]
}

The loop

1. Model responds with tool_calls instead of content.
2. Run the tool yourself, append a role: "tool" message with the result and the matching tool_call_id.
3. Call the endpoint again with the extended conversation; the model produces the final answer.

Cross-provider notes

• Not every model supports tools — the model catalog tags tool-capable models.
• Tool call ids are provider-generated; always echo the id you received, never fabricate one.
• Parallel tool calls (multiple entries in tool_calls) are surfaced as the provider emits them; handle the array, not just index 0.