Self-hosted monitoring for AI coding agents

Why self-hosted matters for agent monitoring

AI coding agents touch your source code. They read files, write code, run tests, and interact with your dev environment. When those agents generate status events, you need to decide where that information goes.

Cloud monitoring services require sending event data to a third-party server. For some teams, that's fine. For others, the idea of agent activity logs leaving the network is a non-starter. The events themselves reveal project names, commit messages, error traces, and workflow patterns. It's operational metadata that can be sensitive.

Radio Agent avoids the question entirely. Everything runs on one machine on your local network. The webhook server, the TTS engine, the audio mixer, and the streaming server all operate locally. The only network traffic is the Icecast audio stream going to devices on your LAN.

What runs where

The entire stack runs as three processes (or three Docker containers) on a single machine:

No component contacts an external server. There are no API keys to configure, no accounts to create, no subscriptions to manage. The install script pulls open-source packages and Docker images, then everything runs locally.

Zero-config install

One command gets you running:

curl -sSL https://radioagent.live/install.sh | bash

The installer detects your platform, pulls Docker images, generates config, and starts all three services. It takes about two minutes. When it finishes, you get a dashboard URL and a stream URL on your LAN IP.

If you prefer to inspect before running:

git clone https://github.com/nmelo/radioagent.git
cd radioagent
docker compose up -d

The result is the same: dashboard at port 8001, audio stream at port 8000.

What you get without a cloud service

• Audio stream with ambient music and voice announcements, accessible from any device on your network (phone, laptop, smart speaker, VLC)
• Web dashboard showing now-playing, a live wire feed of recent announcements with typing animation, and channel controls
• Webhook API that accepts JSON POSTs from any agent, CI system, or script
• Three independent audio channels (music, voice, tones) with per-channel mute and volume
• DJ skill for Claude Code that rewrites robotic templates into natural radio-style speech
• Automatic failure detection through voice and tone differentiation: failures sound different from completions, no scanning required

Comparison with cloud monitoring

Cloud monitoring tools are the right choice when you need long-term retention, cross-team visibility, or integration with incident management workflows. Radio Agent handles a narrower job: real-time, ambient awareness of what your agents are doing right now, from anywhere in the house.

The stack is auditable

Every component is open source:

• Radio Agent (Python, MIT)
• Liquidsoap (OCaml, GPL-2.0)
• Icecast (C, GPL-2.0)
• Kokoro TTS (Python, Apache-2.0)

No proprietary binaries, no phone-home telemetry, no vendor lock-in. You can fork any component, modify the audio pipeline, change the TTS model, or replace Icecast with a different streaming server. The architecture is modular by design: the brain communicates with Liquidsoap via a Unix socket, and with Icecast via standard HTTP. Each piece can be swapped independently.