edytlab
ShippedDescribe an audio edit in English. It renders.
Drop in your tracks and ask for the mashup you want. An agent plans the operations, the DSP runs locally in Rust, and only the chat goes to the model you pick.
- Rust
- Tauri 2
- React 19
- symphonia
- rubato
- Ollama
- commits
- Commits on the main branch of laadtushar/edytlab, counted 10 October 2026.
- tools the agent can call
- The count in docs/tools-reference.md, generated from the tool registry. The agent plans a sequence over them; session state lives in a branchable graph.
The problem
Audio editing has a vocabulary problem. You can hear the edit you want long before you can find the six menu operations that produce it, and that gap is where most people give up.
The approach
A conversational layer over a modular Rust DSP engine. 93 deterministic tools are exposed to the agent, which plans a sequence, and session state lives in a branchable DAG so any plan can be undone or compared against an alternative.
The part that was actually hard
Keeping it local. Decode, routing, effects, resampling and IO are all pure Rust, with the time and pitch work on a phase vocoder written in-house, so the only thing that leaves the machine is the language model request. Stem separation with Demucs and transcription with Whisper are wired in as tools, but their model inference has not shipped yet.
Where it landed
Free and open source under the MIT licence, with installers for macOS, Windows and Linux. The latest release is v0.4.0, of 9 October 2026, and the installers are unsigned, so macOS and Windows warn on first launch. Branchable session graphs, A/B comparison and switching between six model providers all work.