Muse4AI lets AI agents drive Muse over local REST and MCP
A TypeScript app for Windows exposes a signed-in Muse session to ChatGPT, Claude, Cursor, Codex, n8n or any other client that speaks REST, MCP or OAuth.
On 2026-11-14 the GitHub repository andrew-babo/muse4ai listed 783 stars and 64 forks, 11 days after it was created on 2026-11-03. Muse4AI is a local Windows bridge that takes a signed-in Muse session and hands it to AI agents over REST, MCP or OAuth. It is written in TypeScript and released under the MIT license. The project README presents it as running on the user's own PC, with no cloud middleman and no per-seat fee.
What the app does on your machine
Muse4AI runs on Windows and keeps its library, job queue, API keys, tokens and history in %APPDATA%\MuseStudio on the same machine. In the flow the source describes, nothing is relayed through a server the project operates: the agent calls the local app, and the local app talks to Muse using the session already signed in there.
The capabilities it lists:
- Generate images and video, from text alone or with reference photos.
- Chat, where the conversation id comes back so an agent can continue the thread and attach files.
- Extend a finished video, up to 6 steps and about 60 seconds per chain, in the same conversation.
- Edit a video by describing the change in words after attaching an MP4, with no timeline or rendering settings involved.
The app serves its API on http://127.0.0.1:47821. It ships an English and Vietnamese guide with screenshots, and a Skill under agent-kit/ that an agent can read on its own.
Two ways to connect an agent
With an API key: create one in Settings, API & agent (keys start with msk_), then send it as a bearer token. The source says keys are scoped, rate-limited and individually revocable.
With OAuth: the app acts as an OAuth authorization server, publishes /.well-known/oauth-authorization-server, registers clients dynamically and signs its own tokens. You point an agent at http://127.0.0.1:47821/mcp and approve access with the 6-digit code the app shows on screen. Connected apps can be revoked later from the same settings screen.
The MCP tools map onto REST routes: muse_generate and POST /v1/jobs for image or video, muse_chat for a request with kind set to chat, then muse_list_jobs, muse_get_job, muse_cancel_job, muse_extend_video and muse_get_extend. Status sits at GET /v1/status and GET /v1/health, and there is a machine-readable GET /v1/openapi.json plus an llms.txt for agents that prefer Markdown.
The call the source documents
Generating media is one POST to /v1/jobs with the bearer key, a JSON content type and a body naming kind as "video" alongside a prompt string. The app places the job in its queue and the result appears in the library. Repeating a prompt and choosing which account to use are both described as part of the same call.
Install, updates and testing
Grab the installer from the download page and run it; Windows will warn that the file is unsigned, and the source tells users to pick More info then Run anyway. The app then checks for new versions itself and installs them from Settings, About.
To run from source you need Node.js 22 or later, then npm install, npm run build and npm start inside desktop/. Automation runs on every push and every Monday at 02:00 UTC: 848 tests, a strict TypeScript type check, a dependency vulnerability scan, a secret scan across the full git history and live checks against public endpoints. A failed weekly run opens an issue automatically.
Limits and open questions
The project calls itself unofficial and says it is not affiliated with, endorsed or sponsored by Meta Platforms, Inc. It also warns that automating your Muse session can break when the site changes and may not fit Meta's terms of service, which the source tells users to read before relying on the tool. macOS and Linux builds are not packaged or tested, though porting is described as on the roadmap. There is no sign-up, no telemetry server and no subscription.
The source does not explain how the app talks to Muse beyond using your signed-in session, does not give rate-limit figures, and does not say what Muse itself does with generated media or whether accounts can be restricted for automation.
What this suggests for builders
This looks like a template for a pattern rather than a one-off: wrap an existing signed-in web session in a small local server, describe it with MCP and OAuth, and any agent client can use it as a tool. That likely lowers the work needed to give an agent media generation without handing credentials to a third-party cloud or paying per seat.
The trade-off is fragility. Because the app depends on a signed-in session in a web product, changes on that side are likely to surface as breakage, and the terms question the author raises is not one the code can answer. Anyone depending on this for production work should treat availability as outside their control.
For people building with AI
If you are assembling an app that needs image or video output, connecting a model or a service to your app's logic is the part Draftship documents in adding AI to your app. Draftship itself builds apps from plain language and hosts them with a built-in database, user sign-in and hosting, with paid plans adding private projects, a choice of AI model, source code access, custom domains and GitHub push. Muse4AI is an independent project and has no connection to Draftship.
Primary sources
- andrew-babo/muse4ai: Muse4AI — let your AI agent make the pictures and the videos. Local Windows bridge for Muse via REST, MCP, and OAuth. · GitHub trending AI repositories, Nov 14, 2026
This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.