Responses

POST /v1/responses — a stateless OpenAI Responses facade over chat completions.

POST /v1/responses accepts the OpenAI Responses request shape and translates it into a chat completion. It then translates the result back into a Responses response object or Responses SSE events. Routing, quota, sticky conversations, and settlement are the same as for /v1/chat/completions. The facade is stateless. Send the full conversation in input on every call.

Endpoint

POST https://api.malibu.tech/v1/responses
Authorization: Bearer <MALIBU_API_KEY>
Content-Type: application/json

Example

from openai import OpenAI

client = OpenAI(base_url="https://api.malibu.tech/v1", api_key="<your-malibu-api-key>")

resp = client.responses.create(
    model="mlx-community/Llama-3.2-3B-Instruct-4bit",
    instructions="Answer in one sentence.",
    input="What is unified memory on Apple Silicon?",
    max_output_tokens=128,
    store=False,
)
print(resp.output_text)

Response shape. This one is a live 23 Sep 2026 response to the prompt Say hi, trimmed:

{
  "id": "resp_chatcmpl-f3f3409e8f42464ba2bdec2986d0a774",
  "object": "response",
  "model": "mlx-community/Llama-3.2-3B-Instruct-4bit",
  "output": [
    {
      "type": "message",
      "role": "assistant",
      "status": "completed",
      "content": [{"type": "output_text", "text": "Hello!", "annotations": []}]
    }
  ],
  "previous_response_id": null
}

Supported request fields

Field	Translation
`model`	Required. Same IDs as `/v1/models`.
`input`	Required. A string, or a list of message items plus `function_call` / `function_call_output` items.
`instructions`	Becomes a leading `system` message.
`max_output_tokens`	Becomes `max_tokens`. The same 4,096 cap applies.
`stream`	Emits Responses SSE events: `response.created`, `response.output_text.delta`, `response.function_call_arguments.delta`, `response.completed`, and so on.
`tools`	Function tools are passed through. `namespace` tools are flattened into functions. Hosted tools that cannot run on a provider are dropped.
`tool_choice`	Only `auto` takes effect. `required` and a named function tool are accepted but treated as `auto` (not forced). `none` returns `400 unsupported_tool_choice`; omit `tools` instead.
`text.format`	`text`, `json_object`, or `json_schema` (requires `name` and `schema`).
`parallel_tool_calls`, `temperature`, `top_p`, `user`	Passed through.

Rejected

These return 400 with code: "unsupported_parameter" and a param naming the field:

previous_response_id, conversation, and store: true. There is no server-side state, so send the full input.
background: true.
truncation other than "disabled".
response_format. Use text.format instead.

Image, file, or audio input items, and built-in tool call items (web search, file search, computer use, code interpreter, MCP), are also rejected, but with 400 invalid_request and param: null.

Notes

Receipts: the translated body is not an OpenAI chat-completions body, so the receipt verifier’s bundle mode does not apply to it. Use /v1/chat/completions for receipts you plan to verify. Offline verification of current receipts is pending a verifier update. See Signed receipts.
Sticky conversations work the same way. Send X-MacProvider-Conversation. See Sticky conversations.
Errors use the OpenAI error envelope with retryable. See Chat completions → Errors.

Litepaper

Getting started

Buyers

Providers

Network & trust

Guides index

Ledger

Responses

Responses

Endpoint

Example

Supported request fields

Rejected

Notes

​Responses

​Endpoint

​Example

​Supported request fields

​Rejected

​Notes

Responses

Endpoint

Example

Supported request fields

Rejected

Notes