MuseForge

Your own AI media studio,
running on your own GPU.

Video, images, audio, novel-length text and finished audiobooks — one Docker image, no cloud, no per-seat pricing. Use it in the browser, script it over REST, or hand it to an AI agent over MCP.

The Blueprints gallery: reusable recipes labelled by kind
193generation models
184REST endpoints
76MCP tools
33blueprints

More than one clip at a time

MuseForge starts where a prompt box stops: whole stories, whole books, and a voice that stays the same person from chapter one to the end.

Studio

Generate

Video, image and audio models — LTX, Wan, Flux, Qwen, MiniMax H3, SCAIL-2, ACE-Step and more. LoRAs, advanced sampling, multi-clip and extend.

Director

Plan a whole film

Describe a music video or short film. A local LLM plans the shots, writes per-model prompts, renders start frames and runs the pipeline.

Storywriter

Write a novel

Outline, chapters, continuity checks between them, translation into 28 languages, and an audit that reports characters, timeline and plot holes.

Audiobook

Read it aloud

Import a document, assign a voice per passage, add emotion and effects with ducking and loudness matching, render chapters or a chaptered M4B.

Voices

Keep a voice

Build a voice from a description and keep the take you like, or adopt your own recording as a cloning reference. Reused by every book.

Blueprints

Repeat a good result

Save any output's full recipe — model, LoRAs, every setting — and re-apply it in one click. 33 ship with the app across five kinds.

Built to be driven by an agent

Every capability in the UI is a REST endpoint, and a native MCP server exposes 76 of them as tools. An agent can be handed a document and give back a finished audiobook without a human opening the app.

claude mcp add --transport http \
  museforge http://localhost:7861/mcp

Then: list_modelsgeneratejob_statusget_output_url. Or upload_documentaudiobook_createaudiobook_render. Anything without a dedicated tool is reachable through api_request.

Not sure which URL to give it? GET /api/v1/mcp/info answers with the address it was reached at and a ready-made command.

The API & MCP settings panel showing the endpoint URL and client configuration
Settings → API & MCP: endpoint, client config and token status.

Install

An NVIDIA GPU with 6 GB VRAM or more, Docker with the NVIDIA Container Toolkit, and disk space for model weights. That is the whole list.

git clone https://github.com/fgilde/MuseForge.git
cd MuseForge
docker compose up -d

Open http://localhost:7861. Models download on first use, not at install time — the first video can take a while, later runs are fast.

Port 7861 is deliberate. A Maestro instance on the same machine keeps 7860, so both can run side by side. Inside the container the app still listens on 7860.

Full documentation

Where it comes from

MuseForge is a fork of Maestro by @Blizaine, which builds on the Wan2GP generation pipeline. Maestro is a desktop app installed through Pinokio; MuseForge takes the same engine and points it at infrastructure and long-form work. If you want the one-click desktop experience, use Maestro — it is excellent at being that.

Released under the WanGP Non-Commercial Evaluation License 1.1. Free for non-commercial use; your generated outputs are yours.