Tool calling

All three endpoints support function tools in their native format. Other tool types (built-in tools, code interpreter, file search) return 400 unsupported_tool.

The loop is the same everywhere: send tools, the model returns a call, you run it and send the result back.

Chat Completions

{
  "model": "flash",
  "messages": [{ "role": "user", "content": "Weather in Lisbon?" }],
  "tools": [{
    "type": "function",
    "function": {
      "name": "get_weather",
      "description": "Current weather for a city",
      "parameters": { "type": "object", "properties": { "city": { "type": "string" } }, "required": ["city"] }
    }
  }]
}

The reply has choices[0].message.tool_calls. Append that assistant message, then a { "role": "tool", "tool_call_id": "...", "content": "..." } message per call, and send again.

tool_choice accepts "auto", "none", "required" or { "type": "function", "function": { "name": "get_weather" } }. parallel_tool_calls is supported.

Responses

{
  "model": "flash",
  "input": "Weather in Lisbon?",
  "tools": [{
    "type": "function",
    "name": "get_weather",
    "parameters": { "type": "object", "properties": { "city": { "type": "string" } }, "required": ["city"] }
  }]
}

The output contains function_call items. Send them back in input followed by { "type": "function_call_output", "call_id": "...", "output": "..." }. previous_response_id is not supported, so always send the full history.

Messages

{
  "model": "flash",
  "max_tokens": 1024,
  "messages": [{ "role": "user", "content": "Weather in Lisbon?" }],
  "tools": [{
    "name": "get_weather",
    "description": "Current weather for a city",
    "input_schema": { "type": "object", "properties": { "city": { "type": "string" } }, "required": ["city"] }
  }]
}

The reply contains tool_use blocks. Answer with a user message holding tool_result blocks. tool_choice accepts auto, any, none or { "type": "tool", "name": "..." }.