← Open Source
andrew-babo

muse4ai

Muse4AI — let your AI agent make the pictures and the videos. Local Windows bridge for Muse via REST, MCP, and OAuth.

TypeScript
Open on GitHub
Momentum
+122stars in 24 hours+10.0%
1.34k
Stars
73
Forks
+687
This week
2
Contributors
Created 2026-11-03 · Updated 2026-11-19 · #181 today
Top developers
andrew-babo
README

Muse4AI

Let your AI agent make the pictures and the videos.

Muse4AI runs on your own Windows PC and hands your Muse session to ChatGPT, Claude, Cursor, Codex, n8n or anything else that speaks REST, MCP or OAuth — with no cloud middleman and no per-seat fee.

Website · Download · Guide · For agents · Tiếng Việt


How it fits together

   ChatGPT · Claude · Cursor · Codex · n8n · your own script
                       │
                       │  REST  ·  MCP  ·  OAuth
                       ▼
              ┌─────────────────────┐
              │   Muse4AI on your PC│   local library, queue,
              │   (Windows app)     │   keys, tokens, history
              └─────────────────────┘
                       │
                       │  your own signed-in Muse session
                       ▼
                     Muse

Nothing is relayed through a server we run. The app talks to Muse from your machine, with your session, and keeps every result, key and token in %APPDATA%\MuseStudio on that same machine.

What you can ask an agent to do

🎬 Generate video & images

One call, text-only or with reference photos. Repeat a prompt, pick the account, watch it appear in the in-app queue.

💬 Chat that remembers

Ask questions or iterate on an idea. The conversation id comes back, so the agent can keep the thread going and attach files.

⏱ Extend a finished video

Keep a clip rolling, up to 6 steps and about 60 seconds per chain, in the same conversation it was born in.

🛠 Edit a video by describing it

Attach an MP4 and say what to change in words — no timeline, no rendering settings.

🔑 Two ways to connect

An API key you paste once, or real OAuth: the agent asks, and you approve with the 6-digit code the app shows on screen.

📚 Built-in guide

Every screen has a bilingual guide (English / Vietnamese) with real screenshots, plus a Skill your agent can read.

See it in action

Studio — write the prompt, attach references, run. Queue — live status, thumbnails, quick preview.
Library — everything you made, searchable. API & agent — connect an agent in one paste.

Quick start

1. Install (easiest)

Grab the installer from the download page, run it, then add your Muse account in Accounts. Windows will warn that the file is unsigned — choose More info → Run anyway. After that the app checks for new versions by itself and installs them from Settings › About — no need to come back here.

2. Run from source

Needs Node.js 22+.

cd desktop
npm install
npm run build
npm start

The app opens on your desktop and serves its API on http://127.0.0.1:47821.

Connect an agent

Open Settings → API & agent in the app. It shows a ready-made snippet with your key already filled in — paste it into your agent's config.

OAuth (no key to copy)

The app is a real OAuth authorization server: it publishes /.well-known/oauth-authorization-server, registers clients dynamically and signs its own tokens. Point the agent at the MCP endpoint and let it redirect:

{
  "mcpServers": {
    "muse4ai": { "url": "http://127.0.0.1:47821/mcp" }
  }
}

The app then shows a 6-digit code. Type it into the consent page and the token is issued. Anyone who only knows the link cannot grant themselves access. Revoke any app later under Settings → API & agent → Connected apps.

API key

Create a key in the same screen (msk_…), then send it as a bearer token:

curl -X POST http://127.0.0.1:47821/v1/jobs \
  -H "Authorization: Bearer msk_YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"kind":"video","prompt":"a coffee cup on an oak desk, warm morning light"}'

Keys are scoped, rate-limited and individually revocable.

What the agent gets

REST MCP tool
POST /v1/jobs — image / video muse_generate
POST /v1/jobs with kind:"chat" muse_chat
GET /v1/jobs · GET /v1/jobs/{id} muse_list_jobs · muse_get_job
POST /v1/jobs/{id}/cancel muse_cancel_job
POST /v1/assets/{id}/extend · POST /v1/conversations/{id}/extend muse_extend_video · muse_get_extend
GET /v1/status · GET /v1/health muse_status

Full machine-readable contract: GET /v1/openapi.json. For agents that prefer Markdown: llms.txt.

Project layout

Path What lives there
desktop/ The app itself — main process, renderer, REST/MCP/OAuth API, tests, installer
content/ Guide, agent and use-case text shared by the app and the website
agent-kit/ The Muse4AI Skill, so an agent can learn the tool on its own

Development

cd desktop
npm test            # 848 tests
npm run typecheck   # strict TypeScript
npm run package:win # Windows ZIP + installer inputs

Quality checks

Every push, and every Monday at 02:00 UTC, GitHub Actions re-runs the whole suite: 848 tests, a type check, a dependency vulnerability scan, a secret scan across the full git history, and live checks against the public endpoints. If a weekly run fails, an issue is opened automatically.

Honest notes

  • Unofficial. Muse4AI is an independent project. It is not affiliated with, endorsed or sponsored by Meta Platforms, Inc. "Meta" and "Muse" are trademarks of their owners.
  • You own the risk. The app automates your Muse session. Automating a web app can break when the site changes, and may not fit Meta's terms of service — read them before you rely on it.
  • Windows first. macOS and Linux builds are not packaged or tested yet; the core code does not depend on Windows, so porting is on the roadmap.
  • No cloud account. There is no sign-up, no telemetry server and no subscription.

License

MIT © 2026 Andrew Babo. Third-party notices: desktop/LICENSES/.

Questions or ideas: the website · guide · changelog.

Nearby in the ranking

AI Rank for iPhone

Follow AI model rankings

Download on the App Store