September 2026 Why this exists, and where music tools stand with AI agents.
Why Wavelength
Agents already build 3D scenes in Blender and games in Godot. Music had no tool an agent could really pick up, so I built one.
Music was the odd one out.
Claude can drive Blender, Godot and Unreal Engine and make things that are genuinely impressive. I kept waiting for the same thing to happen with music, and it didn't. I wasn't seeing anyone hand a DAW to an agent and get a song back.
Part of that is the tools. A DAW is built around a screen, a mouse and a person listening. An agent has none of those. It can't see the arrangement unless something draws it, and it can't hear the mix at all.
So I stopped trying to fit an agent into a DAW. Wavelength skips the DAW altogether. The agent writes the music as data, and Wavelength plays it through the same synthesizers and sample libraries a producer would use.
Austin Ginder
What an agent needs from a music tool
- No screen. A command it can run, not a window it has to click.
- Files out. A mix and stems on disk, rendered offline, the same every time.
- Sounds by name. "The BBC violins, spiccato", not a knob at 37%.
- Something in place of ears. Loudness, peaks and a picture of the song, so it can judge what it can't hear.
- A person to listen. A way for a human to point at bar 53 and say what's wrong.
DAWs are adding AI, but not for agents.
Every major DAW shipped AI features in the last two years: stem separation, session players, chord detection, assistants. Nearly all of it works inside the app, for the person at the screen. Here is where the big ones stood in September 2026 for an agent working from outside.
| DAW | AI in the app | Can an outside agent drive it? |
|---|---|---|
| Ableton Live 12 | Stem separation (Suite, 12.3), MIDI generators and transformations. | Community MCP bridges built on its undocumented Python remote-script API. Anthropic's Ableton connector answers from the documentation; it can't operate Live. No command-line render. |
| Logic Pro 12 | Session Players, Stem Splitter, Chord ID and a natural-language sound browser. | Community bridges that drive its interface through Accessibility, AppleScript and MIDI. No scripting API for arranging and no command-line render. |
| FL Studio 2026 | Stem separation, AI mastering, and Gopher, an assistant that changes levels, plugin parameters and mixer channels. | Gopher works only inside FL Studio. Scripting is for MIDI controllers and the piano roll; there is a command-line render. |
| Bitwig Studio 6 | None, by choice. | A controller API for transport, clips and device parameters. Its open DAWproject format can be written offline and opened in Bitwig, Cubase and Studio Pro. |
| Cubase 15 | Stem separation (Pro). | A MIDI Remote API meant for hardware controllers. Opens DAWproject files. |
| Pro Tools | Speech-to-text and Splice in the browser. | An official scripting SDK (PTSL) for sessions, editing and export, not for writing music. |
| REAPER | None built in. | The most scriptable: the full ReaScript API and a command-line render. The app still has to run, with a project set up by hand. |
| Suno Studio 2.0 | A generative DAW: songs from text, stems, MIDI and a chat collaborator. | Only through its own chat. No public API. |
From public release notes and reviews gathered on September 26, 2026; each name links to a source. None of these makers offered an official way for an outside AI agent to drive their DAW. The community bridges that do exist steer the app's interface or controller APIs, and they need the app open on a screen. This moves fast: corrections are welcome.
So Wavelength skips the screen.
One command
A JSON job goes in: notes, sounds, effects and the mix. A mastered song, its stems and a report come out. No DAW, no window, on macOS, Linux and Windows.
Real instruments
Any CLAP, VST3 or VST2 instrument on the computer, with presets loaded by name. A built-in synth, drum kit and sampler cover a machine with no plugins at all.
Numbers for what it can't hear
Loudness for every track and section, true peak, checks that name a problem and its fix, and a picture of the whole song for agents that can read images.
A person in the loop
wavelength serve opens a review page on your machine. Pin a comment to bars, tracks or notes; the agent reads it, renders the fix and answers there.
It speaks to agents the way they already work: an agent skill, an MCP server, and the docs built into the binary. Songs stay open in the .wavelength format, with every revision and a stand-in for every sound, so a song made here plays on your computer too.
Where it is today.
Wavelength started on September 24, 2026. 179 commits and 4 releases later, Claude has made every song on this site with it, from chiptune to the BBC Symphony Orchestra.
An agent still can't hear. Wavelength gives it measurements and pictures, and the final call is a person listening. Some plugins won't load without their own window or a licence dialog, and Audio Units aren't supported yet.
Released (v0.4.0)
- CLAP, VST3 and VST2 hosting, Intel-only plugins under Rosetta
- Built-in effects, mastering and loudness targets
- MIDI, MusicXML and DAWproject import, DAWproject export to Bitwig
- The review page,
wavelength serve
Next, in development
- A built-in synth, so a song renders with no plugins
- The song picture and mix checks
- An MCP server and a Claude Code plugin
- The .wavelength format: history, sharing, stand-in sounds