MusicGPT
Rust application that runs Meta's MusicGen locally via ONNX with chat and web interfaces, no Python needed.
About
No Python, no PyTorch, no setup ceremony: MusicGPT packages Meta's MusicGen models into a single Rust binary that runs text-to-music generation locally on macOS, Linux, and Windows through the ONNX runtime. It operates in two modes, a terminal chat interface and a local web UI with chat history, background generation, and on-demand playback, and the web server can be exposed over the network to drive generation from another device. Model sizes from small to large download on first use; the smallest model produces short clips on an Apple Silicon laptop, while larger variants are compute-hungry and benefit from CUDA through the recommended Docker image. Installation comes via Homebrew, cargo install musicgpt, Docker, or prebuilt binaries for each platform. The application code is MIT licensed, though the MusicGen weights it downloads remain under Meta's CC-BY-NC-4.0 non-commercial terms, which limits commercial use of the output pipeline. Around 1,400 GitHub stars reflect its niche as the easiest way to run open music generation models locally, with melody conditioning and infinite audio streams on the roadmap.
Should you use MusicGPT?
Pick it when
Pick MusicGPT when you want to try text-to-music locally without Python or PyTorch: one Rust binary via Homebrew, cargo, Docker, or prebuilt downloads, with a chat or web UI you can reach from another device on your network.
Look elsewhere when
Skip it for commercial output, since the MusicGen weights stay CC-BY-NC-4.0, or for songs with lyrics and vocals, which MusicGen does not do. Larger models are slow without CUDA; ACE-Step covers vocal songs on an 8 GB GPU.
Alternatives to MusicGPT
- AudioCraft
The original MusicGen library with training code, MAGNeT, and JASCO conditioning, more flexible but needs Python 3.9, PyTorch, and ffmpeg.
- Magenta RealTime
Live, steerable generation with Apache code and CC-BY-4.0 weights that permit commercial use, but it works best on Apple Silicon.
- ACE-Step
Generates full songs with vocals and lyrics on an 8 GB GPU, which MusicGen cannot, at the cost of a Python and GPU setup.
- Riffusion
MIT spectrogram diffusion with a Streamlit playground and real-time interpolation, but unmaintained and limited to Python 3.9 to 3.10.
Reviews (0)
Leave a Review
No reviews yet. Be the first to review!
Details
- Category
- Music & Audio Generation
- Price
- Free
- Platform
- Local/Desktop
- Difficulty
- Easy (2/5)
- License
- MIT
- Added
- Jul 29, 2026