Youthe listener and creative director
You give direction in plain language, for example the bass is too busy before the break, make the arp darker, and you accept or reject each audition.
Setloom is an open-source agentic music producer. You direct in plain language and judge by ear. A coding agent composes, renders, and revises the music as editable source.
Keep the ear. Hand the studio to the agent.
Setloom moves music production out of the DAW and into source code that a coding agent writes and renders. You still need a terminal and a frontier model. You no longer need the piano roll, the plug-in chain or the mixing vocabulary. What stays with you is the judgement: hearing a pass and saying what is wrong with it.
Setloom is an open-source harness that lets a coding agent, such as Codex or Claude Code running a frontier model, produce music under a person's direction. The person describes the music and judges each audition by ear. The agent writes each track as editable source (MIDI performances, instrument settings and a renderer), renders it, inspects the result, and revises it. Setloom is made by App Automaton.
Setloom splits music production into three roles: the person directs, the agent produces, and the harness supplies the studio tools.
You give direction in plain language, for example the bass is too busy before the break, make the arp darker, and you accept or reject each audition.
The agent makes the musical choices within the brief, writes each track's source, renders an audition, inspects it, and revises rejected work. Measurements never override your verdict.
The harness provides audio inspection, transcription, stem separation, MIDI and mastering tools, and keeps every render reproducible. It composes nothing.
Each pass goes once round the same six-step loop.
Scroll to go round the loop
Measurements inform the agent. The verdict is always the person's.
Setloom gives a coding agent command-line tools and Python modules for every stage of music production. The agent uses them as each track needs. There is no fixed sequence.
setloom transcribe recovers notes and pitch bends.Basic Pitch, running natively on MLX in FP32setloom transcribe --engine kong recovers piano notes.Kong piano transcriptionsetloom separate writes every stem estimate as float WAV.A 53-stem BS-RoFormer model, converted to run natively on MLXsetloom inspect plots waveform, spectrogram, spectrum, stereo and rhythm, with A/B comparison.NumPy, SciPysetloom.audio measures loudness and applies true-peak limiting.FFmpeg, pyloudnormsetloom play plays a render on request, never automatically.macOS audioSetloom has produced released tracks in electronic, remix and piano work. The published tracks are listed below; all are by App Automaton.
Setloom needs a frontier model running in a coding agent. It is a harness for that agent, not a standalone app, and does not produce music without one.
Setloom installs into one repository-local uv environment in three commands.
# clone the repository and set up its one environment
$ git clone https://github.com/appautomaton/setloom.git
$ cd setloom
$ uv sync --group transcription
# then open your coding agent here and describe the track you hear
The coding agent reads the repository's AGENTS.md and chooses its tools. Full setup is in the README.
Setloom produces editable source that a person revises pass by pass, while prompt-to-song generators such as Suno and Udio return finished audio from a text prompt.
| Setloom | Suno and Udio | |
|---|---|---|
| Output | Editable MIDI, instruments, renderer and mastered audio | A finished audio file |
| Revision | Specific changes to the source, from listening notes | A new generation from a new prompt |
| Who decides | The person, by ear, every pass | The model, per prompt |
| Where it runs | On your Mac, driven by your coding agent | On the vendor's servers |
| Source | Open source, AGPL-3.0 | Closed |