Cua Driver is the UI layer of a computer-use agent that converts model-proposed steps into desktop observations (accessibility tree plus screenshot) and input actions (click, type, scroll, drag) via MCP, CLI, or SDK.

Input delivery follows a background-first ladder—element token, pixel coordinates, browser DOM, then foreground fallback—where background actions never raise windows or move the real pointer. The runtime manages sessions, permissions, and per-action verification across macOS Accessibility, Windows UI Automation, and Linux AT-SPI, reporting whether each action was confirmed, unverifiable, or refused.