Radio Agent is a LAN radio station that gives you ambient audio awareness of your AI coding agents. Music plays continuously. When agents finish tasks, hit errors, or need attention, a voice announces it over the stream.
You hear what's happening from the kitchen, the couch, or wherever you are.
curl -sSfL https://radioagent.live/install.sh | bashOr with Docker:
git clone https://github.com/nmelo/radioagent.git && cd radioagent
docker compose -f deploy/docker/docker-compose.yml upDashboard: http://localhost:8001 | Stream: http://localhost:8000/stream
Task completes -- music ducks, a voice says "eng1 just shipped the auth refactor, single commit, tests green", music fades back up.
Something breaks -- a low tone plays under the music, then the voice: "Build failed on the payments module, missing dependency".
Agent starts working -- a short chime plays under the music. No voice interrupt. You know something kicked off without looking up.
After a day of listening, you stop consciously hearing the tones but you still know when agents are busy. Designed around calm technology principles: information moves between periphery and center, and the system disappears into the background when not needed.
| Channel | What it does | Attention level |
|---|---|---|
| Music | Continuous ambient playback from your library | Peripheral |
| Voice | Spoken announcements via TTS, ducks the music | Center-pull |
| Tones | Short sound effects for agent state changes | Peripheral |
Each channel is independently mutable from the dashboard.
webhook POST -> Brain (FastAPI) -> Kokoro TTS -> WAV file
|
v
Music dir -> Liquidsoap [playlist + crossfade + smooth_add] -> Icecast -> listeners
Two processes. Brain handles webhooks, TTS, and pushes WAV paths to Liquidsoap via Unix socket. Liquidsoap handles all audio mixing, ducking, encoding, and streaming. If Brain crashes, music keeps playing.
Radio Agent ships with a Claude Code skill that transforms robotic agent messages into creative radio callouts. Install it and your agents become DJs.
cp -r skills/dj ~/.claude/skills/djOr download dj.skill directly from the dashboard.
# Simple announcement
curl -X POST http://localhost:8001/announce \
-H 'Content-Type: application/json' \
-d '{"detail":"Phase 1 is complete"}'
# With agent and event kind (triggers voice + tone)
curl -X POST http://localhost:8001/announce \
-H 'Content-Type: application/json' \
-d '{"detail":"Auth refactor done","agent":"eng1","kind":"agent.completed"}'Works with any tool that can POST JSON: Claude Code hooks, GitHub Actions, shell scripts, CI pipelines.
- Python 3.12 with FastAPI/uvicorn
- Kokoro TTS (82M params, 50ms per clip on GPU, works on CPU too)
- Liquidsoap 2.2+ for audio mixing and streaming
- Icecast2 for HTTP audio delivery
Inspired by WRIT-FM, a 24/7 AI-powered internet radio station.
MIT