Tool calling

One schema

Declare tools in OpenAI’s format — tools as an array of { type: "function", function: { name, description, parameters } }, plus an optional tool_choice — and the same body works on every model in the catalogue that advertises the tools capability.

{
  "model": "anthropic/claude-sonnet-5",
  "messages": [{"role": "user", "content": "What is the weather in Lisbon?"}],
  "tools": [
    {
      "type": "function",
      "function": {
        "name": "get_weather",
        "description": "Current conditions for a city.",
        "parameters": {
          "type": "object",
          "properties": { "city": { "type": "string" } },
          "required": ["city"]
        }
      }
    }
  ],
  "tool_choice": "auto"
}

Which models support it is a field on the catalogue, not a guess: capabilities.tools on each row of GET /models.

What translation actually does

For OpenAI, Groq, Together and DeepInfra there is nothing to translate — they all serve /chat/completions with the same body, so the tools block goes upstream verbatim.

Anthropic does not use that format, so the adapter converts in both directions: your function.parameters becomes an input_schema, and the tool_use blocks that come back become OpenAI tool_calls with the arguments re-serialised as a JSON string. From your side the response is the same shape whichever vendor produced it:

All four tool_choice values are mapped, "none" included — it becomes the Messages API’s own {type:"none"} rather than being omitted, which Anthropic would have read as "auto". A named function maps to {type:"tool", name} and "required" to {type:"any"}.

"message": {
  "role": "assistant",
  "content": "",
  "tool_calls": [
    {
      "id": "toolu_…",
      "type": "function",
      "function": { "name": "get_weather", "arguments": "{\"city\":\"Lisbon\"}" }
    }
  ]
}

The loop

Unchanged from any OpenAI client: append the assistant message with its tool_calls, then one role: "tool" message per call carrying the result, then send the whole conversation back. Tool result messages are converted into Anthropic tool_result blocks on the way out, so a transcript built against OpenAI replays correctly against Claude.

Billing and streaming

A turn that only calls tools has no text in it, and it is charged normally — it is the expected shape of a tool-using request, not an empty answer. The zero-completion waiver explicitly checks for tool calls before deciding a response was worthless.

On a stream, tool-call deltas are passed through in whatever shape the provider emitted, so an SDK accumulator assembles them the way it already does.