MazeBench

MazeBench is a local browser app for building and playing persistent 3D puzzle worlds, then evaluating models in the same JavaScript engine.

Quick start

Install from PyPI (Python 3.9+ and Node.js are required):

pip install mazebench
mazebench launch

Or run from a source checkout:

npm ci
npm run dev

The site opens at http://localhost:3000.

Modes

• Play — explore the main world or a local world.
• Build — create and edit worlds stored on your machine.
• Agent — run isolated Prime Intellect Verifiers agents against the maze.

Agent runs

The Agent page runs the stock Verifiers harness in an isolated Prime Sandbox. MazeBench contributes a task-owned Toolset on the evaluator, which advertises only the named game controls. The model receives no shell, filesystem, subprocess, network, repository, hidden state, or scoring capability. Replay video also requires a Chromium-family browser and ffmpeg. Run mazebench --help for commands and options.

For Prime Intellect Verifiers:

pip install "mazebench[prime]"
mazebench prime install
mazebench prime eval model=openai/gpt-5-nano n=1 r=1

Development

npm ci
npm test

Further documentation:

• Prime environment
• Maze level format
• Python packaging

License

MIT