Use case
AI audio editor for DJs
Describe a blend, a longer intro or a loudness target in plain English. An AI agent runs real audio tools on your files, on your computer, and you can A/B every take before you keep one.
A real beatmatch, recorded
This is a screen recording of the desktop app, not a mock-up. The DJ asks for a transition between two tracks and the agent does the rest, ending in an exported file.
Beatmatch and blend two tracks
A DJ asks for both tracks' tempos, and the incoming track is time-stretched to match. It is started 8 bars before the outgoing track ends, the two are crossfaded, and a low-pass filter is put on the outgoing track over the overlap. The mix is compressed, limited at −1 dB, brought to −14 LUFS and exported as a WAV. Along the way Claude points out what it would change: at the 16 s start the bars land a beat apart. Shown at 1.6× speed.
Extend an intro for mixing
A DJ asks for Solar Flare's tempo and bar length, then has the drums-only first 8 bars repeated so the intro runs 16 bars, long enough to mix in over. A high-pass filter goes on the new intro, the first 4 bars fade in, and the track is exported as a WAV. You hear the original intro first, then the new one, then the drop at 0:31. Waits for Claude are sped up; every playback is in real time, with sound.
What you can ask for
Each of these is a tool the agent can call. The full list is in the tools reference.
- Beatmatching. It reads each track's tempo and beat grid, then time-stretches one to match the other without changing its pitch. For a track whose timing drifts, it can warp the audio so the beats land on a grid.
- Key matching. It reads each track's key and can shift pitch in semitones without changing the length. This is most reliable within about a fifth.
- Extended intros and edits. Copy a region, repeat it, paste it elsewhere, cut a range out or move a clip. Because the agent has the beat grid it can talk in bars rather than seconds, as in the recording above, where the incoming track starts 8 bars before the outgoing one ends.
- Transitions. Fades, crossfades, and low-pass or high-pass filters on a track over the overlap, plus EQ, reverb and echo.
- Loudness and delivery. It measures integrated loudness (EBU R128), compresses and limits, brings the mix to a target such as −14 LUFS, and exports WAV, FLAC or MP3.
Try the blend more than one way
Every edit is saved as a node in a branching history. Fork the session, ask for the incoming track to start 4 bars earlier, then A/B the two takes and keep the one you like. Nothing you try is lost, and undo steps back through it.
Your files stay on your computer
The audio processing runs locally in a Rust engine. Tracks are never uploaded. Only the chat goes to the AI provider you pick (Anthropic, OpenAI, OpenRouter, Groq or Gemini), or nowhere at all if you run a local model with Ollama. See the local-first AI audio editor page for how that works.
What it does not do
- It is a prep tool, not a live one. There are no decks, controllers or real-time mixing.
- It cannot split a track into stems or an acapella. Stem separation is not built yet.
- Beat and key detection are software estimates. A track with a drifting tempo or a long ambient intro can be off, and in the recording above the agent itself points out a start where the bars land a beat apart.
- Time-stretch and pitch-shift use a phase vocoder written for the project. It is clean on sustained material, but dense mixes can sound smeared, and the further the change is from 1.0 the more it shows.
- It needs an AI provider: your own key for a hosted one, or Ollama for a local model. The current builds are unsigned, so macOS and Windows warn on first launch; the setup guide has the steps.
Get started
edytlab is free and open source (MIT) for macOS, Windows and Linux. Download it from the releases page, follow the getting started guide, and read the user guide for the other things the agent can do. Posts about how it works are on the blog.