Generate music locally on your Mac.
An open-source desktop AI music generator powered by local inference, built for the OpenMusic series.
| Project | Purpose | Status |
|---|---|---|
| OpenKara | Turn local songs into karaoke tracks with on-device AI stem separation and synced lyrics | Active |
| OpenLoop | Generate new music locally from prompts, lyrics, and musical parameters | Alpha in development |
The shared philosophy is simple: music tools should be local-first, ownership-friendly, transparent, and useful with the media and hardware you already have.
AI music tools are powerful, but many of them share the same problems:
- They require subscriptions.
- They send prompts, lyrics, and creative drafts to cloud services.
- They hide model behavior behind closed platforms.
- They make export, ownership, and reproducibility harder than they should be.
- Text-to-Music Generation — Generate music from prompts such as
lo-fi warm piano, 90 BPM, no vocal. - Lyrics Input — Add lyrics with optional structure tags like
[verse],[chorus], and[bridge]. - Local AI Backend — Run ACE-Step locally through a managed backend process.
- Apple Silicon Acceleration — Use MLX on Apple Silicon with CPU/GPU execution and unified memory.
- Duration Control — Generate clips from short loops to longer song drafts.
- BPM, Key, and Time Signature Controls — Provide musical constraints for generation.
- Seed Reproduction — Reuse a seed to reproduce or iterate on previous results.
- Built-in Preview Player — Play generated audio inside the app.
- Local Generation History — Store prompt, lyrics, model settings, seed, and output path in a local SQLite database.
- Export — Save generated audio to a local output folder.
- Repaint / local audio region regeneration
- Better waveform display
- Multi-model profile management
- More robust model downloader
- Homebrew Cask distribution
- macOS signing and notarization
- Advanced export and audio conversion options
- macOS 14+ recommended
- Apple Silicon Mac recommended
- Node.js 20+
- pnpm 10+
- Rust stable toolchain
- Tauri 2 platform dependencies
- Python 3.11–3.12 for the local ACE-Step backend
uvfor Python environment management
git clone https://github.com/thedavidweng/OpenLoop.git
cd OpenLoop
pnpm install
pnpm tauri devpnpm install
pnpm typecheck
pnpm build
pnpm rust:fmt
pnpm rust:check
cargo test --manifest-path src-tauri/Cargo.tomlDetailed manual QA notes live in docs/testing.md.
For the current implementation status and more development details, see Implementation Status.
| Requirement | v0.1 Target |
|---|---|
| Operating system | macOS 14+ recommended; macOS 12–13 best effort |
| CPU/GPU | Apple Silicon recommended |
| Memory | 8 GB minimum target; 16 GB+ recommended |
| Storage | Several GB for models and generated audio |
| Network | Required for first model/backend setup; offline afterward unless the user chooses otherwise |
Intel Mac support is experimental and outside the v0.1 acceptance target.
OpenLoop uses ACE-Step 1.5 as the local music generation backend.
The app targets a profile-based model setup:
| Profile | Target Device | Default Strategy |
|---|---|---|
| Low Memory | 8 GB Apple Silicon | Conservative settings, lower memory pressure |
| Standard | 16 GB+ Apple Silicon | Recommended default for v0.1 |
| Quality | 24 GB+ Apple Silicon | Higher-quality settings and larger model options |
Model files are downloaded or selected during first setup and stored locally. The application code is MIT licensed; model weights and third-party components follow their upstream licenses.
| Layer | Technology | Purpose |
|---|---|---|
| Desktop framework | Tauri 2 | Rust backend + system WebView desktop shell |
| Frontend | React + TypeScript + Vite | App UI, generation form, player, history panel |
| Backend orchestration | Rust | Process management, API proxy, file operations, SQLite |
| AI backend | ACE-Step 1.5 | Local music generation |
| Apple Silicon inference | MLX | Apple Silicon CPU/GPU execution and unified memory |
| Python environment | uv |
Reproducible local backend environment |
| Database | SQLite | Settings, generation history, backend events |
| Packaging | Tauri bundler | macOS .dmg release |
┌─────────────────────────────────────────────────────────────┐
│ OpenLoop UI │
│ ┌────────────────┐ ┌────────────────┐ ┌───────────────┐ │
│ │ History │ │ Generation │ │ Preview │ │
│ │ Sidebar │ │ Form │ │ Player │ │
│ └────────────────┘ └────────────────┘ └───────────────┘ │
└─────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────┐
│ Tauri Rust Backend │
│ ┌────────────────┐ ┌────────────────┐ ┌───────────────┐ │
│ │ Backend │ │ ACE-Step API │ │ SQLite + │ │
│ │ Manager │ │ Client │ │ File Store │ │
│ └────────────────┘ └────────────────┘ └───────────────┘ │
└─────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────┐
│ Local ACE-Step API Server │
│ Model loading · Task queue · Audio generation │
└─────────────────────────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────┐
│ Local Output Files │
│ WAV / MP3 / FLAC / OGG + generation metadata │
└─────────────────────────────────────────────────────────────┘
OpenLoop uses a local API server model instead of making the UI talk directly to Python. The Rust backend owns process lifecycle, health checks, request validation, file paths, local history, and user-facing error mapping.
OpenLoop is local-first by design.
- Prompts stay on your Mac.
- Lyrics stay on your Mac.
- Generated audio stays on your Mac.
- History is stored in a local SQLite database.
- No account system is planned for v0.1.
- No telemetry is planned for v0.1.
- The app should only use the network for model/backend setup or user-initiated external links.
Logs should avoid storing full lyrics or complete sensitive prompts. Backend errors should be summarized into user-readable messages.
OpenLoop does not provide legal clearance for generated music.
Users are responsible for checking whether generated output is appropriate for publication, monetization, or commercial use. Avoid entering protected lyrics, melodies, voices, or prompts that explicitly imitate protected artists or copyrighted works. When publishing generated music, follow applicable laws and platform rules around AI-generated content disclosure.
- v0.1 targets Apple Silicon first.
- Intel Mac support is experimental.
- First setup may require a large model download.
- Generation speed depends heavily on memory, model profile, duration, and inference settings.
- Repaint is planned after the first Alpha.
- The app does not guarantee copyright-free output.
- The current UI favors local workflow coverage and technical completeness over final visual polish.
Contributions are welcome once the initial Alpha structure is in place.
Recommended contribution areas:
- macOS packaging
- Tauri backend process management
- ACE-Step API integration
- generation history UX
- model setup diagnostics
- low-memory performance testing
- documentation
Before opening a large PR, please open an issue describing the proposed change.
OpenLoop application code is released under the MIT License.
Third-party models, libraries, and tools retain their own licenses. In particular, ACE-Step, MLX, FFmpeg, Tauri, and other dependencies should be reviewed according to their upstream license terms before redistribution.
OpenLoop builds on work from the open-source music and local AI ecosystem, including:
- ACE-Step 1.5
- MLX
- Tauri
- The broader open-source audio tooling community
OpenLoop is part of the OpenMusic series by David Weng.
For the detailed development roadmap and planning documents, see:
- Implementation Status — Current implementation status and completed milestones
- Development Plan — Full development timeline, phases, and technical planning
- Testing Guide — QA procedures and manual test coverage notes