by oil-oil
AI-powered Werewolf (Mafia) social deduction game where every player is controlled by top LLMs like DeepSeek, Qwen, Gemini, and more
# Add to your Claude Code skills
git clone https://github.com/oil-oil/wolfcha
Current sponsors:
Note: This project was born at the "Watcha + ModelScope Global Hackathon" as an AI-native game.
"Wolfcha" combines Wolf (Werewolf) + Cha (猹, a character from Chinese literature). It's a nod to the hackathon host while also capturing the fun of watching AI characters interact — like spectating a show.
After graduating, getting 8-12 people together for a proper Werewolf game became nearly impossible. While Werewolf is fundamentally a social game, its core appeal — logical deduction, verbal sparring, and reading between the lines — remains captivating even without the social element.
To enjoy Werewolf anytime, anywhere, we built this AI-powered version. As the name suggests, every player except you (Seer, Witch, Hunter, Guard, Werewolves, etc.) is controlled by AI.
Thanks to the growing context windows and instruction-following capabilities of large language models (LLMs), we've implemented a sophisticated dual-layer roleplay system:
<div align="center"> <img src="https://img.shields.io/badge/DeepSeek-V3.2-1B75FF?style=for-the-badge" alt="DeepSeek" /> <img src="https://img.shields.io/badge/Qwen-Qwen3-5A6CFF?style=for-the-badge" alt="Qwen" /> <img src="https://img.shields.io/badge/Moonshot%20AI-Kimi-111111?style=for-the-badge" alt="Kimi" /> <br/> <img src="https://img.shields.io/badge/ByteDance-Seed-333333?style=for-the-badge" alt="Seed...No comments yet. Be the first to share your thoughts!
Every conversation is generated in real-time, full of unpredictability and fun.
This isn't just Werewolf — it's a battleground for AI models.
We've integrated multiple top AI models to compete against each other. After each game, you can see which model powered each character. It's a hidden "Turing test" — observe which model reasons sharply, which one seems "adorably clueless," and which one sounds the most human.
Current built-in models (subject to project configuration):