Werewolf AI

The project began as a question rather than a product: whether a large language model can play a social deduction game for real — bluffing, reading the table, holding a secret across a long night — instead of merely following the rules. Werewolf, also known as Mafia, was chosen because it punishes everything an LLM finds difficult. The author's conclusion after many rounds is that the strongest models now do manage it: they form alliances, make accusations under pressure, defend themselves, and occasionally land a bluff.

At the table each bot is given a hidden role, its own backstory, a play style and a voice, and the site is explicit that the bots do not know which of them are AI. Rounds are themed — the examples run from a haunted manor to a submarine crew — with illustrations and voice acting, so the same line-up does not play out the same way twice. Putting providers in one game is the point rather than a detail: the head-to-head is what makes a round interesting, and the same idea is formalised in a benchmark section that invites you to judge which model is the best liar and which one you can fool.

Play happens in the browser after signing in with Google or GitHub. The free tier asks for no card and charges nothing: a stated $5 of AI usage a day up to $20 a month, up to five games a day, on a curated, price-banded model set, with voice acting and illustrations included. Paid is deliberately not a subscription — you pre-load a balance and pay only for what you use, which opens the full model catalog, removes the daily and per-game limits, and adds unlimited portrait rerolls and night replays, billed at the provider's base price plus a stated 15% margin. You never supply your own API keys; every tier runs on the site's own provider accounts.

The write-up is unusually candid about what was hard. Keeping a model in character across a long game meant fighting context rot, invented facts and forgotten roles, and the author describes that as the real engineering work rather than the game design. The source is published on GitHub without a license file, so it can be read but is not licensed for reuse, and development conversation happens in a public Discord.

Playing a full Werewolf round solo against bots, comparing two or more providers head-to-head on the same table, and using the benchmark rounds as a practical read on which model keeps a lie standing longest.

The policy, last updated October 14, 2026, states that sign-in uses Google and GitHub OAuth and stores your public name, email address and profile picture; that game messages, actions and usage statistics are kept in Firebase Firestore to run the game and track spending; that card payments are handled entirely by Stripe and never seen by the operator; and that game messages are sent to the AI providers to generate bot responses. Application logging runs through BetterStack with a three-day retention. It also states that player data is not used for anything beyond making the game work.

Official privacy policy