ChatArena: Multi-Agent Language Game Environments for LLMs

Introduction

ChatArena (or Chat Arena) is a library that provides multi-agent language game environments for Large Language Models (LLMs) like GPT-3, GPT-4, and ChatGPT. The primary goal of ChatArena is to facilitate research into autonomous LLM agents and their social interactions, helping to develop their communication and collaboration capabilities.

It offers several key features:

Abstraction: A flexible framework for defining multiple players, environments, and their interactions, based on Markov Decision Process.
Language Game Environments: A collection of environments to aid in understanding, benchmarking, or training agent LLMs.
User-friendly Interfaces: Both Web UI and CLI are provided to develop and prompt engineer LLM agents within these environments.

Important Note: As of August 11, 2025, this project has been deprecated due to a lack of widespread community use and is no longer planned to receive any additional updates or support.

Installation

To get started with ChatArena, ensure you have Python >= 3.7. An OpenAI API key is optional if you plan to use GPT-3.5-turbo or GPT-4 as an LLM agent.

Install with pip:

pip install chatarena

Alternatively, install from source:

pip install git+https://github.com/chatarena/chatarena

To use GPT models, set your OpenAI API key:

export OPENAI_API_KEY="your_api_key_here"

Optional dependencies can be installed for full functionality:

pip install chatarena[all_backends] # for all supported backends
pip install chatarena[all_envs]     # for all environments
pip install chatarena[all]          # for full functionality

You can also launch a demo Web UI locally:

pip install chatarena[gradio]
git clone https://github.com/chatarena/chatarena.git
cd chatarena
gradio app.py

Examples

ChatArena provides a variety of language game environments:

Conversation: A multi-player environment simulating a conversation, such as the "NLP Classroom" example.
Moderator Conversation: Based on conversation, but with a moderator controlling game dynamics. Examples include "Rock-paper-scissors" and "Tic-tac-toe".
Chameleon: A multi-player social deduction game where players describe clues about a secret word, and a "chameleon" tries to blend in without knowing the word.
PettingZooChess: A two-player chess game environment integrated with PettingZoo.
PettingZoo TicTacToe: A two-player tic-tac-toe game environment driven by hard-coded rules, distinct from the moderator-driven version.

Why Use It

ChatArena was designed to be a valuable tool for researchers and developers interested in the capabilities of Large Language Models in multi-agent settings. Its structured approach allows for:

Research and Development: A dedicated platform for exploring and benchmarking LLM agents' communication and collaboration skills.
Flexible Experimentation: The abstract framework enables easy definition and customization of new language games, players, and interaction rules.
Diverse Environments: A suite of pre-built environments, from simple conversations to complex social deduction games and classic board games, provides immediate testing grounds.
User-Friendly Interaction: Both command-line and web-based interfaces simplify the process of developing and prompt engineering LLM agents.

While the project is now deprecated, its design principles and the environments it offered provide insights into the challenges and opportunities in multi-agent LLM research.