Cua: Open-Source Computer-Use Infrastructure & Drivers for AI Agents

The AI coding landscape is transforming, with AI assistants now handling graphical user interface tasks. This shift is known as Computer-Use 2.0, where AI agents dynamically switch between shell scripts, APIs, and graphical applications. Developers building these agents face challenges with brittle graphical drivers, lack of display servers in headless environments, and interference with physical mouse control.

Cua is an open-source infrastructure platform designed to address these issues. It acts as the operating system and automation layer for any AI agent or model. Cua comprises five modules: the Cua Driver for background desktop automation, Cua Fleets for ephemeral cloud desktops, CUA-S1 for fast decision models, Lume for Apple Silicon VM management, and Cua Bench for standardized evaluation.

Cua Driver integrates natively across macOS, Windows, and Linux, allowing background desktop interactions without disrupting the user's workflow. Cua Fleets provide isolated, sandboxed Linux desktops for untrusted agent execution or scalable evaluation runs. CUA-S1 offers compact decision models designed for rapid UI element decisions, while Lume manages local macOS and Linux virtual machines.

Finally, Cua Bench offers a framework for defining desktop tasks, running reference solutions, verifying task completion, and exporting trajectories for model training. By providing cross-platform drivers, sandboxed environments, and specialized models, Cua enables AI agents to fully leverage human computing capabilities, representing a significant advancement for the agent ecosystem.