Muse MuseConcept
Muse MuseConcept

Independent concept built on Meta's Model API. Not an official Meta product.

How it works

Tell it an idea. Muse Muse builds a working app on Meta's AI, tests it and puts it online at its own link.

Every method below is written up in the Muse Muse docs, Meta's Model API and Muse Code documentation in one place.

Five steps, a few minutes

  1. 1. Share an idea

    Type it, or drop in photos, screenshots, a sketch, a short video, a voice memo or a PDF.

    Files are stored at unguessable links, so anyone with the link can open them. Voice memos are transcribed with Muse Voice Transcribe, and images, video and PDFs go to the model as they are.

    In the docs:Image understandingVideo understandingSpeech to textFile handling

  2. 2. Get a plan

    Spark 1.3 Contributor turns it into a short plan, called a brief.

    The brief comes back as strict JSON: the screens, what goes in and comes out, which Meta models the app calls and why, honest limits and a cost per use. If an idea can't be built safely, the brief says so and the build stops there.

    In the docs:Structured outputModelsReasoning

  3. 3. Muse Code builds it in a sandbox

    Meta's coding agent writes the app in its own sealed computer while you watch.

    Each build gets its own Vercel Sandbox: a small Linux virtual machine started from a snapshot that already has a starter app and Muse Code installed. Muse Code runs headless (muse exec --json --yolo) with the brief as its prompt, AGENTS.md as the project rules and a skill for each Meta model. The firewall blocks everything but npm, Meta's API and Muse Muse. Muse Code only ever holds a stand-in key, and the firewall swaps in the real one on the way out. Every step streams into the build room next to a live preview.

    In the docs:Muse CodeHeadless modePermissions and the sandboxAGENTS.md and settingsSkills, hooks and subagentsContained execution

  4. 4. It tests and fixes itself

    The app is checked like a real one, and anything broken goes back to Muse Code.

    In the same sandbox: a typecheck, a production build, the app started for real, and a browser test that clicks through the main flow on a phone-size screen. Then Spark 1.3 Contributor judges screenshots of every screen. Problems go back to the same Muse Code session (resumed with --session-id), for up to 3 fix rounds.

    In the docs:Launch flagsBrowser-verified web designJudging screenshots

  5. 5. It goes live

    It gets its own link, a demo video and a spot in the gallery.

    Each app is deployed as its own project. It calls Meta's models only through Muse Muse's gateway, with its own token and a daily spending cap, so no API key is ever in the app. Its actions become an MCP connector that the Muse app and other AI tools can call.

    In the docs:AuthenticationPricing and rate limitsTool calling

What you get

  1. Its own link

    A working app anyone can open on a phone or a computer.

  2. A demo video

    A short recording of the app doing its job, ready to share.

  3. Use it from the Muse app

    Each app becomes a connector, so the Muse app and other AI tools can use it for you.

  4. Keep improving it

    Ask for changes in the build room and it builds a new version.

Good to know

No sign-up and no API keys. Most builds take a few minutes. Apps reach Meta's models only through Muse Muse, which keeps the keys and caps what each app can spend. See the gallery for apps people made, Rebuilds for Meta's older demos rebuilt on today's models, and Usage for every Meta token used.

Make something