---
name: meta-muse
description: Write and reason with Meta's Muse Wonder (1M-token context) and make images with Muse Image through Meta's Model API. Use when the user wants Muse, needs to feed a very long document, video or audio file to a model, or wants a cheap image. Needs META_MUSE_API_KEY.
---

# Meta Muse

Key: `META_MUSE_API_KEY`. Get one at https://dev.meta.ai/platform under API keys.

Examples pipe JSON through `jq`. Without it, read the JSON and take the field named in each step.

## Write something

The API speaks Anthropic's Messages format at `https://api.meta.ai`:

```bash
curl -s https://api.meta.ai/v1/messages \
  -H "Authorization: Bearer $META_MUSE_API_KEY" -H "Content-Type: application/json" \
  -d '{"model":"muse-wonder-1.0","max_tokens":4000,"messages":[{"role":"user","content":"Outline a 10-episode podcast about street food."}]}' \
  | jq -r '.content[] | select(.type=="text") | .text'
```

Muse thinks first, and the thinking counts toward `max_tokens`, so leave room. The reply starts with a `redacted_thinking` block; the answer is the block whose `type` is `text`.

OpenAI's Responses format works too: `POST https://api.meta.ai/v1/responses` with `{"model":"muse-wonder-1.0","input":"..."}`.

## Make an image

```bash
curl -s https://api.meta.ai/v1/images/generations \
  -H "Authorization: Bearer $META_MUSE_API_KEY" -H "Content-Type: application/json" \
  -d '{"model":"muse-image-1.0","prompt":"Isometric pixel art of a tiny record shop"}' \
  | jq -r '.data[0].b64_json' | base64 --decode > image.webp
```

The picture comes back as base64 WebP in `data[0].b64_json`.

## Notes

- Models: `muse-wonder-1.0` (text, reads images, video, audio and PDFs), `muse-image-1.0` (images, about a cent each), `muse-voice-transcribe-1.0` (speech to text).
- Muse Wonder is in preview and isn't offered in every country.
- The Messages endpoint rejects `stop_sequences`, `top_k` and `thinking: {"type":"disabled"}`. Leave them out.
- Bro can run Claude Code and Codex on Muse: pick "Meta (Muse)" as the provider.