Steady State Radio ← back to the stream

A radio station
that runs itself.

This page documents the whole machine — the models, the agents, the guardrails, and what it actually costs. No mystique: the interesting part is that it's all quite simple.

The premise

Most "AI radio" is a playlist with a chatbot name. Steady State Radio is an experiment in autonomous media: the music is generated in-house with an open-source model, the programming follows the clock, operations are handled by agents, and the station heals itself when something breaks. One human sets the rules, holds the budget, and can stop everything with a single command.

The question being tested: how much of a real media property can run in steady state — continuously, cheaply, and honestly — with humans only at the edges?

The stack
Sound
ACE-Step 1.5 → catalog
Open-source music model, commercial use permitted. Tracks are generated in batches, by style family and daypart. Marginal cost per track: effectively zero.
Quality gate
Quarantine → automated QA → air
No track reaches rotation directly. Every batch lands in quarantine, where a QA script checks duration, loudness (−16 LUFS target) and — after one track slipped through repeating the same piano note — a loop detector built on windowed RMS autocorrelation.
Broadcast
AzuraCast AutoDJ, 24/7
Scheduled dayparts (First Light, Deep Block, Afterglow, Night Shift) over a general catalog, with a fallback playlist if everything else fails.
Operations
n8n agents
A watchdog checks the stream every 5 minutes and restarts the station via a locked-down Docker proxy if it dies. A kill switch pauses all paid generation. A daily report posts uptime, listeners and spend.
Voice
Vega — an AI, and says so
Station IDs and (soon) short spoken segments, synthesized locally with Kokoro TTS. Disclosure is a house rule, not a footnote.
Supervision
One human, three levers
Budget cap, content rules, kill switch. That's the whole job description.
Proof it heals itself
// incident log · 2026-07-03 · live test

The station's own container was deliberately killed as a live test. The watchdog detected the dead stream, restarted the broadcast through its Docker proxy, verified the signal, and posted a report — back on air in 181 seconds, no human involved. That test is now part of the deployment checklist: an ops agent that hasn't survived a real failure is decoration.

What it costs (really)

The entire station runs on one small VPS that also hosts other projects. Approximate monthly numbers, in the open:

Server share (VPS, streaming + agents)~$5–8
Music generation (ACE-Step, batched)~$0–2
Agent LLM calls (small models, capped)~$1–3
Domain~$1
Steady state≈ $10/mo

// hard cap US$15/month, enforced by an agent that pauses generation — never the stream — if spend crosses the line.

The charter, summarized
  • All music is AI-generated with openly licensed models. No human artists imitated on purpose.
  • The AI never pretends to be human. Vega discloses what she is, on air and on this site.
  • No health claims, no sleep/therapy language, no urgency, no hype. Focus music, plainly.
  • The stream never stops for budget reasons — only non-essential generation pauses.
  • Every sponsorship is answered by agents but co-signed by the human. One sponsor per segment.
The human

Steady State Radio is built and supervised by one person as part of a broader experiment in agent-operated products. The interesting failures and the architecture are documented as the station evolves — the machine is the product, but the notes are the point.

Listen to the result →