A Framework for LLM Social Games That Never Found Its Audience

What it does

ChatArena is a Python framework that drops multiple LLM agents into structured language games—tic-tac-toe, rock-paper-scissors, social deduction, and free-form conversation—then manages their turns, observations, and rewards through a Markov Decision Process abstraction. It wraps environments, language backends, and player policies into a single loop, offering both Web UI and CLI frontends for running or debugging agent interactions.

The interesting bit

The framework treats natural-language observations as the only channel between agents and the hidden game state: players never see raw state vectors, only prose descriptions. This forces LLMs to reason and strategize purely through text, making the setup a useful probe for emergent social behavior.

Key highlights

• Ships with ready-made environments including a social-deduction “Chameleon” game, chess via PettingZoo, and moderator-driven turn-taking games.
• Supports swapping LLM backends—OpenAI GPT, Anthropic, Cohere, Hugging Face—without rewriting environment code.
• Provides both a Gradio Web UI and an interactive CLI for prompt-engineering agents and watching them interact.
• Exposes an Arena class that drives the main loop, plus hooks for custom environments, backends, and player policies.

Caveats

• The project was deprecated in August 2025 because of a “lack of wide spread community use” and will receive no further updates or support.
• The README cites an arXiv paper “coming soon” since at least 2023; it is unclear whether it was ever published.

Verdict

Worth a look if you are maintaining legacy multi-agent LLM experiments or want a reference architecture for language-only game environments. Skip it if you need an actively supported framework—this codebase is frozen.