Initialize the environment

A suite of 100+ single-, two-, and multi-player text-based games for benchmarking and training LLMs.

Play | Leaderboard | Games | Examples

Introduction

TextArena is an open-source framework for evaluating and training language-model agents through competitive and cooperative text-based games.

It provides 100+ single-, two-, and multi-player environments through an interface styled after OpenAI Gym, ranging from classic board and card games to negotiation, social deduction, and multi-agent reasoning tasks. TextArena makes it easy to plug in language-model agents, run self-play or model-vs-model evaluations, and build training pipelines around interactive environments. It also supports 192 languages, enabling the same environments to be presented to agents through different language interfaces.

Updates

• 14/08/2026 TextArena goes multilingual: 192 languages are now supported.
• 28/05/2026 We released the findings of MindGames.
• 31/03/2026 We released the MindGames NeurIPS 2025 game trajectories.
• 26/11/2025 We added negotiation games.
• 31/07/2025 We added SettlersOfCatan to TextArena.
• 14/07/2025 We announced MindGames, a NeurIPS 2025 competition for training LLMs on TextArena games that require theory of mind.
• 01/07/2025 We released v0.6.9 with 100 games, simplified states, new observation wrappers for training, and default environment wrappers.
• 01/07/2025 We released SPIRAL: Self-Play on Zero-Sum Games Incentivizes Reasoning via Multi-Agent Multi-Turn Reinforcement Learning, introducing RL through self-play on TextArena games as a potential new training paradigm.
• 22/06/2025 We released UnstableBaselines, a lightweight asynchronous online RL library for training LLMs on TextArena games.
• 16/04/2025 We released the TextArena paper.
• 14/02/2025 We released the new stable version on both PyPI and the website.
• 31/01/2025 We released the initial demo, which was highlighted by Andrej Karpathy — and promptly crashed all our servers.

Getting Started

Installation

Install TextArena directly from PyPI:

pip install textarena

Offline Play

Agents only need to implement a __call__ function that accepts a string observation and returns a string action. We provide several basic agents here.

The example below lets GPT-4o-mini play against anthropic/claude-3.5-haiku in a game of TicTacToe.

We use the OpenRouterAgent, so first set your OpenRouter API key:

export OPENROUTER_API_KEY="YOUR_OPENROUTER_API_KEY"

Then initialize the agents and environment:

import textarena as ta

# Initialize agents
agents = {
    0: ta.agents.OpenRouterAgent(model_name="GPT-4o-mini"),
    1: ta.agents.OpenRouterAgent(model_name="anthropic/claude-3.5-haiku"),
}

# Initialize the environment
env = ta.make(env_id="TicTacToe-v0")

# Optional visualization wrapper
env = ta.wrappers.SimpleRenderWrapper(env=env)

env.reset(num_players=len(agents))

done = False
while not done:
    player_id, observation = env.get_observation()
    action = agents[player_id](observation)
    done, step_info = env.step(action=action)

rewards, game_info = env.close()

Multilingual Support

TextArena supports UI localization across 192 languages for more than 60% of its games. See textarena/envs/README.md for a complete list of games with multilingual support, and textarena/utils/locales for the localization files covering all 192 supported languages. Languages can be assigned independently to each player using lang_mapping, allowing the same environment to be presented through different language interfaces.

Examples of multilingual game interfaces:

Language Coverage

• 8 languages are manually reviewed by native speakers.
• 42 additional high- and mid-resource languages are translated and automatically verified.
• 142 low-resource languages are produced using open machine translation and evaluated using a multi-model fidelity pipeline.

Low-resource localizations are machine-verified rather than native-reviewed and are labeled by confidence tier.

Multilingual Usage

Languages can be assigned independently to each player using lang_mapping in env.reset():

import textarena as ta

agents = {
    0: ta.agents.OpenRouterAgent(model_name="GPT-4o-mini"),
    1: ta.agents.OpenRouterAgent(model_name="anthropic/claude-3.5-haiku"),
}

env = ta.make(env_id="TicTacToe-v0")

env.reset(
    num_players=len(agents),
    lang_mapping={0: "en", 1: "de"},
)

done = False
while not done:
    player_id, observation = env.get_observation()
    action = agents[player_id](observation)
    done, step_info = env.step(action=action)

rewards, game_info = env.close()

from textarena.utils.locales.language_confidence import warn_if_flagged

warn_if_flagged(lang)

Citation

If you use TextArena in your research, please cite:

@misc{guertler2025textarena,
    title={TextArena},
    author={Leon Guertler and Bobby Cheng and Simon Yu and Bo Liu and Leshem Choshen and Cheston Tan},
    year={2025},
    eprint={2504.11442},
    archivePrefix={arXiv},
    primaryClass={cs.CL},
    url={https://arxiv.org/abs/2504.11442},
}

Contributing

All forms of contribution are very welcome. Whether you're adding new games, improving existing functionality, fixing bugs, or helping with documentation and translations, we'd be glad to have your help.

Check out the open issues or join us on Discord to get started.

Some examples:

• Make RushHour board generation algorithmic.
• Extend FifteenPuzzle to arbitrary sizes.
• Review multilingual translations.
• Improve rendering, tests, or tooling.