Two players join a room. An AI plays the host and the rest of the suspects — each with a secret motive. Every case is generated live, so it never plays out the same way twice. Shipped as a real product: bilingual, voiced, monetized.
A short look at the product: a case generated from nothing, two players at the table, and the moment the accusation lands.
Scripted mystery kits are great — once you've bought one, scheduled six people, and read the host booklet for an hour. The format barely exists for two people on a weeknight, and once you've solved a script, it's spent.
Most mystery games fall apart below 4–6 players. Couples and remote pairs are left out.
A good session needs a game master who knows the truth, paces the reveals, and never breaks. That person can't really play.
Once you know the culprit, the box is dead. Fixed content doesn't replay.
Pick a script and the system generates a complete, self-consistent, locked case behind the scenes — truth, culprit, motive, timeline, evidence and red herrings — then runs it as a live multiplayer session.
For mystery scripts, every character — including the real players — gets a plausible motive and a secret. Only one is the real killer; the rest are deliberate misdirection. The truth is locked at generation and can't drift.
AI characters act on their own goals: they accuse, defend, form alliances, and actively try to frame others — including you. They never speak for the real players, and they're influenced by what's actually said at the table.
Your friend clicks a link and joins instantly — no install, no signup. Private role info is isolated per player; public scene is shared and synced live.
Optional neural TTS reads narration and each character in a distinct, gender-matched voice. At the end, every player gets a multi-metric scorecard on deduction, roleplay and key calls.
The shared scene streams to both players in real time; your private role and secret stay yours alone. AI suspects push their own agendas — here one of them is openly trying to frame you.
Every turn returns strict JSON (narration, NPC lines, revealed evidence, private notes, act state) with schema enforcement, retry, and an OpenAI fallback when the primary model filters or truncates.
The full case is generated once and stored server-side only; the host model is fed the secret file but is hard-blocked from leaking it, from spoon-feeding clues, and from impersonating real players (prompt + server-side filtering).
5–8 AI-designed acts advance on a live timer (or host control), gated so you can only skip ahead once the act's key clue is actually found. Clues only surface on genuine investigation, never on autopilot.
Supabase Realtime drives instant updates, with a polling fallback so invited (anonymous) players never desync — a bug I only caught by playing on two devices.
One endpoint, swappable between OpenAI and Azure neural voices by env flag, with per-character voice selection by gender and hash, a sequential playback queue, and storage caching so a line is never re-synthesized.
Stripe credit packs for hosting, Google OAuth, a per-account credit ledger with webhook dedup, a free first game for every new account, and invited guests always free.
Pick a script and the engine generates a complete, self-consistent case server-side: culprit, motive, timeline, evidence and deliberate red herrings. It's scored for solvability, then frozen — the host model is fed the secret file but hard-blocked from leaking it.
Custom domain, auto HTTPS, bilingual UI that auto-detects the browser language, link-preview cards, and a private live-analytics panel for the operator.
Design, full-stack build, LLM orchestration, voice, payments, auth, deploy, and growth — one person, end to end.
When a player commits to an accusation, the locked truth is revealed and the case is walked back act by act. Each player receives a multi-metric scorecard on deduction, roleplay and the key calls they made.