Cua gives AI agents isolated desktop control

Restricting AI agents to fragile web scrapers and bespoke API wrappers limits real-world computer use. Moving agent execution into dedicated, isolated virtual machines solves both user disruption and security risks, though VM latency and sandbox management remain active engineering hurdles.

• –Background execution: Cua interacts directly with native OS accessibility trees, allowing agents to execute tasks in the background without stealing user mouse or keyboard focus.
• –Lightweight virtualization: Integration with Apple's Virtualization Framework through Lume provides near-native execution performance on Apple Silicon without heavy container emulation overhead.
• –Universal agent connectivity: Support for the Model Context Protocol (MCP) and typed SDKs allows existing coding agents like Claude Code and Cursor to operate full desktop software seamlessly.
• –Standardized evaluation: Cua Bench provides reproducible sandboxed testbeds to benchmark computer-use agents and train specialized small decision models like CUA-S1.