Muse API quickstart
Use your Quotaflow API key with the request format your client already supports. Start with Chat Completions for a new text integration.
Model IDs and access
| Model | Public model ID |
|---|---|
| Muse Wonder 1.0 | meta/muse-wonder-1.0 |
| Muse Wonder Full 1.0 | meta/muse-wonder-full-1.0 |
| Muse Spark 1.3 | meta/muse-spark-1.3 |
These are catalog identities, not a promise that every model or capability is available to every key. First get a Quotaflow API key, then check access using the same key your application will use:
export QUOTAFLOW_API_KEY="YOUR_QUOTAFLOW_API_KEY"
curl https://api.quotaflow.ai/openai/v1/models \
-H "Authorization: Bearer $QUOTAFLOW_API_KEY"
Find the exact meta/muse-* ID in data[].id. Use that ID in the examples below. If it is missing, check your key's Muse access before integrating. Listing a model does not guarantee that every optional request shape has available capacity.
Choose an endpoint
| Client format | URL | Output limit | Read the answer |
|---|---|---|---|
| OpenAI Chat Completions | https://api.quotaflow.ai/openai/v1/chat/completions | max_completion_tokens | choices[0].message.content |
| OpenAI Responses | https://api.quotaflow.ai/openai/v1/responses | max_output_tokens | output message items; SDK: output_text |
| Anthropic Messages | https://api.quotaflow.ai/v1/messages | max_tokens | content blocks with type: "text" |
The OpenAI SDK base URL is https://api.quotaflow.ai/openai/v1. An Anthropic-compatible client uses https://api.quotaflow.ai as its base URL. Do not append /openai/v1 to the Messages URL.
Chat Completions: cURL
curl https://api.quotaflow.ai/openai/v1/chat/completions \
-H "Authorization: Bearer $QUOTAFLOW_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "meta/muse-wonder-1.0",
"messages": [
{"role": "user", "content": "Design an order-management API and explain its main endpoints."}
],
"max_completion_tokens": 4096,
"stream": false
}'
Read choices[0].message.content. For a text conversation, include earlier user and assistant messages in the next request's messages array. Tool conversations must also preserve the tool calls and matching results.
Chat Completions: Python
Install the OpenAI Python SDK with pip install openai, then set QUOTAFLOW_API_KEY in your environment.
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["QUOTAFLOW_API_KEY"],
base_url="https://api.quotaflow.ai/openai/v1",
)
response = client.chat.completions.create(
model="meta/muse-wonder-1.0",
messages=[
{"role": "user", "content": "Design an order-management API and explain its main endpoints."}
],
max_completion_tokens=4096,
)
print(response.choices[0].message.content)
For streaming, set stream=True and read each chunk's choices[0].delta.content when present; some chunks contain only control or usage information. Streaming availability depends on the selected model and request shape. See Chat Completions.
Stateless Responses
curl https://api.quotaflow.ai/openai/v1/responses \
-H "Authorization: Bearer $QUOTAFLOW_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "meta/muse-wonder-1.0",
"input": "Design an order-management API and explain its main endpoints.",
"max_output_tokens": 4096,
"store": false,
"stream": false
}'
Read text from output items of type message, whose content includes output_text blocks; do not assume the first output item is a message. The OpenAI SDK provides response.output_text as a convenience.
Muse Responses is stateless: Quotaflow keeps no response store, so send the full conversation on each request. To continue a conversation, put the previous response's output items back into input unchanged, including reasoning and function-call items, append any matching tool results, and send include: ["reasoning.encrypted_content"] with store: false, the stateless way Muse documents for carrying reasoning between turns.
Do not use previous_response_id, conversation, prompt, background, or store: true. Meta's own Muse API offers stored history and background mode; Quotaflow's Muse Responses does not, by design. A request with a non-null previous_response_id or conversation, any prompt, or background: true is answered 400 invalid_request_error with code: "unsupported_parameter" and param naming the field. See Responses API.
Anthropic-compatible Messages
Use this format only when your client expects Anthropic request and response shapes and your key has Muse access. Muse remains a Muse model; using this endpoint does not make it a Claude model or enable every Claude-specific feature.
curl https://api.quotaflow.ai/v1/messages \
-H "x-api-key: $QUOTAFLOW_API_KEY" \
-H "anthropic-version: 2023-06-01" \
-H "Content-Type: application/json" \
-d '{
"model": "meta/muse-wonder-1.0",
"max_tokens": 4096,
"messages": [
{"role": "user", "content": "Design an order-management API and explain its main endpoints."}
],
"stream": false
}'
Read text blocks from the response's content array; it may also contain thinking or tool-use blocks. Messages streaming uses Anthropic events, not Chat Completions chunks.
- Discover Muse with
GET /openai/v1/models. The AnthropicGET /v1/modelslisting remains Claude-only and is not a Muse access check. POST /v1/messages/count_tokensdoes not support Muse and returns404for a Muse model.- Start without explicit thinking or sampling overrides. For Wonder 1.0, do not send
thinking: {"type": "disabled"},tool_choice: {"type": "none"}, ortop_p: 0. If using client tools, let the model choose automatically rather than forcing a named tool ortool_choice: {"type": "any"}. - Use only capabilities enabled for the selected model and request. A Messages capacity gap returns
529 overloaded_error; it does not mean the model ID should be replaced with a Claude ID. - On
meta/muse-wonder-1.0,stop_sequencesworks: when a text block of the answer contains a requested sequence, Quotaflow ends the text before it, drops what follows, and returnsstop_reason: "stop_sequence"with the matchedstop_sequence. Quotaflow searches each text block on its own, only before the answer's first tool call, tool result or citation, and only when the request cannot produce citations (it has no tool other than a client tool, no MCP server, and no document or search result with citations enabled). An answer outside that scope is returned as generated. - Muse prompt caching is automatic. A
cache_controlbreakpoint withttl5mor1his accepted under the same breakpoint rules as any Messages request (for example, at most four breakpoints, and no1hbreakpoint after a5mone) so Anthropic clients work unchanged, but it does not set a cache lifetime: Muse caching has no5mor1hsetting, so the value you choose does not change how Muse caches. meta/muse-wonder-1.0does not currently return thinking content on Messages, even withthinking: {"type": "adaptive", "display": "summarized"}.
Output budgets, errors, and billing
Meta documents max as the default reasoning effort of meta/muse-wonder-1.0 and high as that of meta/muse-spark-1.3, used when a request sets none. On Responses, Quotaflow sends the reasoning.effort you set unchanged.
Muse models reason by default. An output budget can be consumed by reasoning before visible answer text appears. If the response reaches its output limit before producing the answer, increase the endpoint's output-token budget within the model's supported limit. An empty text field alone does not prove success or failure: inspect completion status, stop reason, and usage.
401: check the Quotaflow key and authentication header.403: check the key's Muse entitlement.400: check the endpoint's field names and unsupported parameters; do not retry the same invalid request unchanged.429,503, or Messages529: use bounded retries with backoff for temporary capacity errors. Avoid replaying a stream after partial output without handling possible duplicate work.
Prices and billing terms depend on the exact model and your account. Check the dashboard and billing documentation; this guide does not quote a universal rate or guarantee live model capacity. Keep API keys on your server, not in browser code.
Text Muse is not Muse Voice
meta/muse-wonder-1.0 does not accept video input content blocks on Messages: a request carrying one is answered 400 invalid_request_error before it reaches the model, and it is not billed. Other optional media support must be checked per model; a catalog identity is not a media-capability promise. Audio file transcription uses the separate Muse Voice endpoint and its own model ID, WAV format and request fields, not this text API.