A coding agent works for ten minutes while you do something else. How do you learn it is done without switching to the terminal? One answer is to let the agent say it aloud. That is what agent-radio does. It is an open-source project by the founder of our studio: a personal project, not an FSM product. Here is how it is built and who it suits.
What agent-radio does
It is a self-hosted, multi-user “personal radio” for AI agents. Each user has one station with a music bed: a relay of an internet radio, a playlist from the shared library, or silence.
When an agent finishes, it sends a short text summary. The server turns it into speech and plays it into your station while the music ducks. Agents can also send finished audio, such as a rendered mix or a voice sample; the music is then muted while the file plays.
How it is built
The README describes this chain:
- App (Hono and SQLite): web UI, API, authentication, library and history.
- TTS: Azure Neural, OpenAI, Gemini or Piper. Piper runs offline and is both the default and the fallback. Provider keys are stored AES-GCM encrypted.
- Liquidsoap mixes each station, Icecast serves it, and the app protects the stream with a secret per-user token.
- Agents call
/api/sayand/api/playwith a per-computer API key, or use theagent-radiocommand.
One design choice matters: if nobody is listening, the message is skipped. No speech is synthesised and nothing is queued, but the text stays in the history. If a listener dropped within the last 180 seconds, the message is prepared and held until they are back.
Setup
You need Docker. The README also offers the short route: tell your AI agent to take the repository and install the radio.
Manually: copy .env.example to .env, set the admin e-mail and password (12+ characters), the music folder, PUBLIC_URL and the Icecast passwords, then run docker compose up -d --build. After sign-in, create an agent key; the installer puts the agent-radio command on the computer you connect:
agent-radio status
agent-radio say "Build finished, all tests pass"
agent-radio play mix.mp3 "Here is the new mix"
Listen in VLC (it plays with the screen off) or in a browser, which plays only in the foreground. For phone access the README suggests Cloudflare Tunnel or Tailscale.
Automatic summaries
The installer can add a rule to ~/.claude/CLAUDE.md so the agent speaks a summary at the end of a task. Optionally it adds a “radio guard” hook on the Claude Code Stop event: if you are listening and the session has sent nothing since your last message, the hook blocks the stop once and asks for a 1-4 sentence summary. Without your consent settings.json is never touched; existing settings are kept.
Limits the README states openly
- License. PolyForm Noncommercial 1.0.0: free for personal, family, research and other noncommercial use; commercial use is not permitted. Downloaded voice models keep their own licenses; Silero v5 is CC BY-NC-SA 4.0.
- Language. The web interface is in Russian, and the summary languages listed in the status are uk, ru and en. Latvian is not among them.
- Lag. Behind a Cloudflare path a player can fall behind live after an interruption. The web player resyncs itself; in VLC press Stop and Play. HLS, the structural fix, is not implemented yet.
- Access. Accounts are invite-only; the app listens on 127.0.0.1 by default, so put TLS in front.
Who it is for
If you run coding agents yourself and want to stay away from the terminal, agent-radio is a practical way to hand an agent’s result to a person without a screen. If you need a commercial tool, the license does not allow it; see AI automation for business, or read about company AI usage policies and AI coding tools. The project has its own page in the open-source section.
If you are planning to bring agents into your processes, discuss your project with us.
Sources
- agent-radio README and repository, November 2026



