A friend in another city opens a link on their phone. A map loads — a candlelit tavern, their character token standing by the door. They type what they want to do. A few seconds later a voice narrates the scene back to them: the smell of the hearth, the innkeeper looking up, a die already rolling behind the words.
The game master running that game — narrating, rolling, tracking the world, handling combat, keeping four players straight — is not a service I pay per message for. It’s an AI daemon — one of a few I keep running — on a MacBook on my desk. And each turn of that game costs, to a very good approximation, zero dollars.
This post is how that works, and why the “zero” is the interesting part.
Why it’s free
Most AI products are metered. You call a model API, you pay per token, and a chatty application becomes a running bill. A live game — where every player action is a fresh generation of narration, dice resolution, and world updates — would burn tokens continuously.
The trick here is that the brain isn’t a metered API at all. It’s a persistent AI process running under a flat-rate plan I already pay for — the same subscription I use for everyday work. From the game’s perspective, generating a turn of narration costs the same as the process sitting idle: nothing marginal. The meter isn’t running because there is no meter.
Everything around the brain is chosen to stay at zero too:
- The public URL players hit is fronted by a free tunnel — no domain-hosting bill, no load balancer.
- The always-on piece that holds game state and serves the web client is a tiny cloud box small enough to live in a free or near-free tier, and it sleeps when nobody’s playing.
- The connection from that box back to the laptop is a private encrypted tunnel — also free.
So the whole cost structure is: a subscription I’d pay anyway, plus infrastructure that rounds to zero. The marginal cost of one more turn, one more session, one more player, is nothing.
The path a turn takes
Here’s the trip a single player action makes, start to finish:
- Phone → public URL. The player’s browser talks to a normal web address, fronted by the free tunnel.
- → the cloud box. A small always-on relay serves the game client and holds the authoritative game state — the map, the tokens, whose turn it is.
- → private tunnel → the laptop. When a turn needs a decision — narration, a skill check, a combat resolution — the relay hands it across an encrypted tunnel to the Mac.
- → the daemon. A persistent AI process reads the turn: the scene, the character, the action. It writes the narration, resolves the dice, updates the world.
- → all the way back. The result flows back out to the relay, out through the tunnel, and onto the player’s screen — as text, and as spoken audio.
No player ever knows a laptop is involved. From their phone it’s just a game that answers.
The actually-hard part
The economics are a trick. The engineering is a real problem, and it’s this: a persistent, interactive AI process is not naturally a service.
The daemon I’m using is built to be talked to — an open-ended assistant, not a request/response endpoint. Turning it into something a game can call, one turn at a time, reliably, meant building a bridge: a small message channel the game posts a turn into, and the daemon answers back through, in a structured shape the game can render. The brain proposes the narration and the dice; the machinery around it enforces the turn order, the map, the rules.
Getting a headless, always-on AI to behave like a dependable backend — to wake for a request, answer in a fixed format, and go quiet — was most of the work. The model was never the bottleneck. The plumbing to make a conversational thing act like a service was.
Once that bridge existed, the rest of a real game followed: a consistent narrator voice, real dice a player can roll from the screen, turn-based combat on a rendered map, multiple players sharing one game master, and text-to-speech so the story is heard, not just read.
Why it matters
The lesson isn’t “run your production app on a laptop.” It’s a hobby rig — one brain, it naps when idle, and there are rough edges I’m still smoothing. Don’t put your business on it.
The lesson is about where the cost of an AI product actually lives. For most builders it’s the model bill, and that bill is what makes ambitious ideas — a game that generates every scene live, an assistant that never sleeps — feel financially impossible for one person. But if the brain runs on a flat plan you’re already holding, that number collapses. What’s left is just plumbing, and plumbing you can learn.
Flip that one assumption and the space of things a single person can build gets a lot bigger. A friend playing a live, AI-narrated game from their phone, for free, is a small proof of a larger point:
Your laptop can be the brain behind a real-time app people actually use. The model was never the expensive part. It was just the part everyone else is still paying for.