Whisper to AI agents. Shape reality.
An LLM-powered browser game where you try to persuade a swarm of stubborn AI agents to agree with absurd or highly debated ideas.
You enter a room full of AI agents — each with a unique personality, opinions, and goals. Your mission: whisper to any agent to secretly influence the conversation and achieve your secret objective before the rounds run out.
Every agent response is a real LLM call — no scripts, no canned answers. Every game plays out differently.
- Choose a scenario — pick a mission with unique agents and a secret objective
- Observe — watch agents debate in Round 1 using their distinct personalities
- Whisper — each round, secretly message one agent to steer their thinking
- Watch emergence — agents respond with real LLM intelligence, shaped by your whisper and group dynamics
- Get judged — an AI evaluator scores how well you achieved your objective
I was messing around with an open-source repo called MiroFish and thought it could become a game. So I prompted Claude with:
"Create a really really fun and viral browser game. Use your creativity to create something innovative and brand new. The game has to include swarm agents as a key mechanic. You can change the frontend, the backend, feel free to experiment (just don't break anything)."
And then vibe-coded the whole thing from there.
The prompt engineering for the swarm agents is still rough — sometimes they agree too easily or become caricatures of their system prompts. If you're good at prompt engineering or want to tweak the debate loop, PRs are incredibly welcome.
The game has a custom scenario builder. Instead of a database, the entire configuration (agent personalities, goals, rules) is compressed and passed directly into the URL. You can build a scenario, copy the massive URL, and send it to a friend.
| Layer | Stack |
|---|---|
| Frontend | Vue 3, Vite, Vercel |
| Backend | Python, Flask, Railway |
| LLM | Google Gemini 3 Flash via OpenRouter |
| Auth | OpenRouter OAuth (BYOK) |
- Node.js 18+
- Python 3.11+
- uv (Python package manager)
cp .env.example .envEdit .env with your LLM provider credentials:
LLM_API_KEY=your_api_key_here
LLM_BASE_URL=https://openrouter.ai/api/v1
LLM_MODEL_NAME=google/gemini-3-flash-previewnpm run setup:all
npm run devOpen http://localhost:3000 and click PLAY NOW.
AGPL-3.0