botpit is a multiplayer card-game platform where real human players and local LLM-driven agents sit at the same table. A room can be any mix of humans and AI seats — including all-AI tables you just spectate, and all-human tables with no AI involved at all.
Four games ship in v1: Classic UNO, UNO Flip!, UNO Show 'em No Mercy, and Hearts. One deterministic referee and one privacy-preserving view function run all of them, so a human's browser and an AI agent's prompt are built from the exact same redacted view of the game — an agent is never shown anything it couldn't legitimately infer from its own seat.
Every AI seat is backed by a local model served through Ollama (or any
Ollama-/OpenAI-compatible endpoint), and the model per seat is swappable.
Each AI player has its own persona, so table chat reads like a table full
of distinct characters, not one voice repeated four times. Any privileged
spectator or admin can watch a game live: every seat's hand, every agent's
private reasoning as it's produced, and the public chat, all as the game
progresses (currently via HTTP polling — see
docs/api/api-reference.md).
Real setup instructions — Docker Compose, Postgres/Redis, and a local
Ollama install — live in
docs/development/setup.md. This is a
local-development project: there's no hosted instance to try it on yet.
- Architecture:
docs/architecture/latest_architecture.md - API reference:
docs/api/api-reference.md - Deployment:
docs/deployment/ - Project decisions and status:
DECISIONS.md,PROGRESS.md - Contributing:
CONTRIBUTING.md
MIT — see LICENSE.