Noir Casefile

A murder mystery game where AI writes a new case every time you play.

What is it?

Someone has been murdered. You're the detective.

1. Read the case. You get a victim, a crime scene, some clues and a few suspects.
2. Ask questions. Talk to the suspects. You only get a few questions, shared between all of them, so choose carefully.
3. Find the lie. One suspect is the killer, and their story has a hole in it.
4. Make your accusation. Name the killer, the weapon and the clue that proves it. Get all three right to win.

Every case is made up fresh by AI, so no two games are the same. Your game saves automatically, so you can close the tab and come back later.

Screenshots

Desktop

Mobile

Difficulty levels

Level	Suspects	Questions	The killer is...
Rookie	3	12	nervous and slips up easily
Detective	3	10	evasive about details
Hardboiled	4	10	calm, with a red herring suspect to fool you
Noir	5	8	ice-cold, with several red herrings

Run it yourself

You need Node.js and a free Groq API key.

git clone https://github.com/denzeltl/noir-casefile.git
cd noir-casefile
npm install
cp .env.example .env.local

Open .env.local and paste your key:

GROQ_API_KEY=your_key_here

Then start it:

npm run dev

Go to http://localhost:3000 and open a case.

Tip: If you hit Groq's free rate limit, add a second key as GROQ_API_KEY_BACKUP. The game switches to it automatically.

Ad Studio

Behind /admin there's a small studio that turns finished cases into 15-second vertical trailers. You pick which cases get an ad. AI writes and scores the hooks, you approve them, your PC renders the video, and you post it by hand.

The graph

brief ─▶ hookWriter ─▶ critic ─┬─▶ pickHook ─▶ storyboardWriter ─▶ render ─▶ approveVideo ─▶ done
              ▲                │   (you)                            (PC)       (you)
              └── best < 7/10 ─┘      │ rewrite                        ▲            │ re-render
              (max 3 drafts)          └──▶ hookWriter                  └────────────┘

• hookWriter (Groq) writes five hooks from the spoiler-free brief. The full case, with the killer and the solution, never enters the graph.
• critic (Gemini) scores them out of 10. If the best is under 7, the writer gets the critique and tries again, up to three drafts.
• pickHook pauses for you: pick a hook, edit it, or ask for a rewrite with a note.
• storyboardWriter turns the hook into on-screen text for the trailer template.
• render pauses until the render worker on your PC has made the MP4 with HyperFrames and uploaded it.
• approveVideo pauses for you again: approve it, re-render it or discard it.

Each run page shows a timeline of the steps it went through, with times and the models used. A failed run says which step broke, and Retry run runs that step again without redoing the work before it. A run that stops making progress for 10 minutes is marked as stalled and can be retried the same way.

Why LangGraph

• Loops. The writer and critic loop is a conditional edge with a hard limit, not hand-written retry code.
• Durable pauses. The two human decisions and the render handoff are interrupt()s. The run's state is checkpointed in Postgres, so a run can wait for days, survive a server restart, and resume exactly where it stopped. The render worker on your PC resumes it the same way the admin does.
• Two providers. Each step names a primary and a fallback model. When Groq or Gemini hits a free-tier rate limit, that step switches to the other provider instead of failing the run.

Run it

You also need a Supabase project and a Gemini API key.

1. Run supabase/schema.sql in the Supabase SQL editor.
2. Add the Ad Studio variables from .env.example to .env.local: NEXT_PUBLIC_SUPABASE_URL, SUPABASE_SERVICE_ROLE_KEY, DATABASE_URL, GOOGLE_API_KEY, ADMIN_PASSWORD, N8N_SHARED_SECRET and AD_STUDIO_URL.
3. Run npm run dev and open http://localhost:3000/admin. Every case a player opens appears there.
4. To render videos, keep npm run render-worker running on your PC.
5. Optional: Telegram alerts and a Create ad button in Telegram. See n8n/README.md.

To try the graph without the admin, npm run ad-graph runs it once on the newest saved case (or npm run ad-graph -- <caseId>), in memory.

Built with

Next.js, React, Tailwind CSS, Motion, and Groq (qwen/qwen3.8-27b) for the AI.

License

MIT. See LICENSE.