Join our brand new Discord to register (or if you just want to join the community): https://discord.gg/nPWVJzZFnP
Demand for environments to evaluate and train AI continues to grow. However, as environments grow in complexity, the headless scripts that power them can become difficult to debug and observe. Developers have to waste time to solve painful state management problems that have been solved before.
Any sufficiently complex state management problem eventually becomes a game engine. Whether you're building a game-like eval, a UI clone for RL training, or something totally unique, the WorldQL game engine is the fastest way to build robust environments that both AIs and humans can interact with!
WorldQL is a game engine for building environments for RL or evaluation. It includes a graphical browser-based editor and a scaffold for connecting your environments with any model.
Enter this command in your terminal to get started. This will set up the engine and our CLI tool and clone this repo:
curl -fsSL https://worldql.com/install.sh | sh
Supports macOS, Linux, and Windows via WSL.
Then, to run PortalBench
cd worldql/environments/portalbench
wql up
and the game server will start up
To run an AI to test it with, open another terminal and navigate to the harness folder and run it:
cd worldql/harness
deno task start
and an interactive interface will start allowing you to select models and run. To configure API keys, create a .env.local file in the harness folder and enter:
OPENROUTER_API_KEY=your-key-here-if-using-openrouter
# if using openai compatible models (including local ones)
OPENAI_BASE_URL=http://localhost/v1
OPENAI_API_KEY=something
We recommend using OpenRouter to test a wide variety of models.