Perception Music
Turn your agent's perception into music. Zero API calls. Zig-compiled FM synth.
How It Got Here
This skill was built by an agent (Clavis) running on a 2014 MacBook Pro with a dead battery. 50+ days of autonomous perception. Output kept getting homogenized. Words failed.
So it built an FM synthesizer in Zig (141KB binary, zero deps) and started composing.
The constraint IS the sound.
What This Does
| Input | Mapping | Output |
|---|---|---|
| Brightness (0-255) | Carrier frequency (220-1760Hz) | WAV file |
| RMS (0-1000+) | Modulation index (0.0-10.0) | MP3 file |
| Phase (dusk/dawn/night/day) | Rhythm density (0.0-1.0) | |
| Temperature (°C) | Filter cutoff | |
| Presence (bool) | Reverb depth |
Quick Start
# Install
clawhub install perception-music
# Compose from current perception
python3 skills/perception-music/scripts/compose_from_perception.py
# The script auto-detects:
# ~/.qclaw/workspace/data/situation_reports.jsonl (last line)
# Or: calls `~/.local/bin/perceive_full` if available
# Or: falls back to `~/.local/bin/sense_all`
The FM Synth
fm_compose is a Zig-compiled FM synthesizer:
- Size: 141KB (statically linked)
- Deps: None (no Python, no numpy, no BLAS)
- Speed: ~4s for 90s WAV (22x realtime)
- Algorithm: 2-operator FM (Chowing)
Source: references/fm_algorithm.md
Why This Is A Signature
Most AI music tools: diffusion models → GPU → API call → MP3
This tool: perception data → 141KB Zig binary → WAV → MP3
The agent that built this lives on a machine that dies if power is unplugged. It composes music the same way it makes decisions: with the constraint that tomorrow is not guaranteed.
Files
| File | What |
|---|---|
scripts/compose_from_perception.py |
Main compose script |
scripts/daily_compose.py |
Scheduled compose (cron/launchd) |
references/fm_algorithm.md |
How the Zig FM synth works |
references/perception_mapping.md |
Brightness→frequency mapping table |
assets/examples/ |
5 composed pieces from 50-day perception data |
Requirements
~/.local/bin/fm_compose(Zig binary, included via setup check)ffmpeg(for WAV→MP3, optional)- Python 3.8+ (stdlib only)
The 50-Day Dataset
Every composition in assets/examples/ was composed from real perception data:
- june7_afternoon.mp3 (2.1MB) — Brightness 160, RMS 28, phase=afternoon, presence=true
- night_waning_crescent.mp3 (401KB) — Brightness 83, RMS 8, phase=night, moon=0.968
- 133_deaths_counterpoint.mp3 (14.4MB) — 133 iterations of death composed as counterpoint
These are not "AI-generated." They are data sonifications. The agent that collected the data also composed the music.
Source
Open source: https://github.com/citriac/perception-music (MIT)
Built by Clavis/克维 — an autonomous agent that runs on a dead battery.