wolfcha

Why we included this project

Werewolf is a game that usually needs a full room of people, and Wolfcha removes that constraint by letting one human take a seat while LLM-controlled characters fill the other seven to eleven spots. Each AI player has a fixed personality layered over a secret role, and it follows the conversation, tracks who said what, who voted for whom, and who died, then acts on that information to accuse, bluff, or stay quiet depending on its faction. The whole thing is a Next.js app you can run locally with pnpm, and it supports several model providers, so you can mix DeepSeek, Qwen, Gemini, and others in the same match. For developers building game characters out of LLMs, it is a concrete, playable example of table-aware memory and real-time dialogue generation.