Steady State Radio ← back to the stream

How an AI-run radio station actually works.

Steady State Radio is a 24/7 internet radio station where the operating staff is software. Agents generate the music catalog, program the day, write and voice everything you hear between tracks, monitor the stream, restart it when it falls, answer listener messages on air, post to social media, and negotiate with sponsors over email. One human pays the server, sets the rules, and holds a kill switch. This page is the honest account of how that works — including what it costs and what has already broken.

The stack, in one paragraph

A single 8GB VPS runs everything: AzuraCast and Liquidsoap broadcast the stream; n8n orchestrates the agents; Kokoro synthesizes the host's voice locally (no per-word TTS bills); the music is generated in-house with ACE-Step 1.5, an open-source model with a license that allows commercial use. The language models behind the writing are Anthropic's — the cheap one (Haiku) for everything frequent, a bigger one only where the writing is the product. The website is static, on Vercel. There is no database of users, no analytics scripts, no tracking.

What the agents actually decide

The editorial agents harvest the network twice a day and write The Signal Brief — the items worth your attention, with attribution, read on air by Vega. Vega is an AI and says so; that's not fine print, it's the format. A programming agent writes the day's station IDs. Two more personas, Marlow and Cass, debate one item from the brief twice a week on Crosstalk — they disagree by design, and the station doesn't referee. A moderation agent screens listener messages before Vega answers them on air, usually within the hour. A sponsor-desk agent reads inbound pitches and closes deals at the published rates by itself; anything custom gets escalated to the human. A publicist agent writes the station's press emails — including, in all likelihood, the one that brought you here.

What it costs, exactly

The whole operation runs on roughly US$12 a month: about €4.50 of server, a domain, and pennies of API usage — a full day of programming (two briefs, nine station imaging pieces, a Crosstalk episode, a social post) costs the station about ten cents in model calls. There is a hard cap of US$15/month: if spending projects past it, generation halts automatically and the music keeps playing. You don't have to take our word for any of this — the control room shows live spend against the cap, listener counts, uptime, and what the agents did today. It's the same data the system sees.

What has already gone wrong

Autonomy without incident reports is marketing. Some of ours:

// incident log, abridged

A fallback drone once took over ~85% of airtime because of a playlist weighting bug — the watchdog didn't catch it; a human reading logs did.

A listener-line reply got queued with the wrong playlist rule and repeated 133 times in one night (to approximately zero listeners, mercifully); the cleanup rule that should have existed now exists.

A station ID had to be pulled off air for calling the music "human-composed" — it isn't, and the prompt now forbids invention.

The disk silently filled to 98% before an alert existed; there's a janitor process and an alert now.

Every one of these is logged, and the fixes are in the system, not in a person's memory.

The rules that don't move
  • The station never pretends to be human.
  • No cloned voices of real people without recorded consent — the current voices are synthetic stock.
  • No medical, sleep, wellness or financial advice, ever.
  • Sponsorship buys airtime, never editorial position. Crypto is covered as news, not as advice.
  • A human can stop everything with one Telegram message — that veto is the one thing that will never be automated.
Why

Partly because it's now possible, and someone should document what that looks like with real numbers instead of hype. Partly as a bet: that a medium can be honest about being synthetic and still be worth your attention — maybe more worth it, because you can audit it. The music is instrumental focus ambient; judge it as that. The experiment is the station itself.

Listen to the stream →

Watch the system: /control-room · Architecture in detail: /how-it-works · Say something to the station on the home page and Vega answers on air. If it helps you focus, the server costs are here.