Tool calling, JSON mode and vision

Both wires accept their own native shapes for tools, structured output and images, on every model. Where a model's own protocol differs from the wire you called, Sator translates the request in and the response out; where they match, the request passes through untouched. Either way you write the request in the dialect of the endpoint you called.

Tools#

	OpenAI wire	Anthropic wire
Definition	`tools[].function` with `name`, `description`, `parameters` (JSON Schema), optional `strict`	`tools[]` with `name`, `description`, `input_schema`
Choice	`"auto"`, `"none"`, `"required"`, or `{"type": "function", "function": {"name": "…"}}`	`{"type": "auto"}`, `{"type": "any"}`, `{"type": "none"}`, or `{"type": "tool", "name": "…"}` (the strings `"auto"`, `"none"`, `"required"` also work)
Parallel calls	`parallel_tool_calls: false` to disable	`tool_choice.disable_parallel_tool_use: true`
Results	`role: "tool"` message with `tool_call_id`	`tool_result` block naming `tool_use_id`

json

{
  "model": "deepseek-v4-flash",
  "messages": [{"role": "user", "content": "What is the weather in Montreal?"}],
  "tools": [{
    "type": "function",
    "function": {
      "name": "get_weather",
      "description": "Current weather for a city",
      "parameters": {"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"]}
    }
  }],
  "tool_choice": "auto"
}

When streaming, tool-call deltas are forwarded exactly as the model emits them.

Structured output#

OpenAI wire	Anthropic wire
`response_format`: `{"type": "text"}`, `{"type": "json_object"}`, or `{"type": "json_schema", "json_schema": {"name", "schema", "description"}}`	`output_config.format`: `{"type": "json_schema", "schema": {…}, "name"}` — `json_schema` is the only format accepted

Vision and documents#

OpenAI wire	Anthropic wire
`image_url` content parts, base64 data URLs only (`data:image/png;base64,…`). Remote URLs are not fetched.	`image` and `document` blocks with `source.type: "base64"` and `source.media_type`, or `source.type: "url"`

Images count toward the 5 MB body cap.

Reasoning#

OpenAI wire: reasoning_effort is passed through; models that expose their reasoning return it as reasoning_content on the assistant message, and you may send it back on prior assistant turns.
Anthropic wire: thinking blocks are accepted in the conversation; redacted_thinking blocks are dropped silently, the one deliberate silent drop on either wire.

What is refused#

An unsupported value is a 400 invalid_request_error naming the field — for example an unknown tool_choice string, a response_format type other than the three above, a content part of an unknown type, or a message role the wire does not define. Nothing is silently discarded except redacted_thinking. The full list of statuses is in Errors.