meirtz/shinken

Shinken is an open-source project for researchers who want to train AI agents to control real desktop computers. Think of it as a training gym where the AI learns to click, type, and navigate a computer screen the same way a person would, by looking at pixels rather than through special shortcuts built into individual apps. The goal is to make that training process faster and cheaper at scale. The core problem Shinken solves is repetition. When training an AI agent on computer tasks, you typically have to reset the desktop to a known state thousands of times per experiment. Normally that means rebooting, reinstalling software, or re-navigating to a specific point in a workflow, which wastes time. Shinken takes a snapshot of a running desktop at any moment, then creates copies of that exact state in roughly 0.1 to 0.6 seconds each. You reach a state once and fork it as many times as you need, rather than rebuilding it from scratch every time. The scale claims are specific and backed by benchmarks the authors say are rerunnable. According to the README, a single laptop can hold over 8,000 live desktop environments at once on a single processing thread, and 128 fully booted real desktops can be launched in about 7 seconds. Each step where the agent acts on the desktop and takes a new screenshot costs around 13 milliseconds. For developers who want to connect their own code, Shinken provides a Python interface. You create a session, and that session gives you a live desktop where you can call functions like click, type_text, and screenshot. The same interface is designed to work across Docker-based local desktops and other backends. The project is in early stages. The README is direct about which parts are working today under automated tests and which parts are still design. It uses Docker and Python 3.10 or above, and is licensed under Apache 2.0. The full README is longer than what was shown.