Summary

Background operation is the whole point — the machine stays yours while the agent works.

Most desktop automation takes over the pointer, which means you stop working while it runs. Cua Drivers act without stealing cursor or focus, and Linux is documented with explicit limits on raw background input rather than a blanket claim. If what you need is a disposable machine rather than your own, the sandbox side of the same project is the other half.

What it is

The MCP server that ships with Cua Drivers, Cua's background computer-use layer. Agents click, type and verify inside native desktop applications while you keep using the machine: the driver does not take the cursor or the window focus. The same CLI and MCP server work on macOS, Windows and Linux, with X11 and compositor-specific Wayland routes on Linux.

What you get

Native desktop apps driven in the background from Claude Code, Cursor, Codex, OpenClaw or your own client
One toolset across macOS, Windows and Linux, rather than a per-platform automation stack
The wider Cua stack alongside it: agent-ready sandboxes for Linux containers and VMs, macOS, Windows and Android through one API, cloud or local via QEMU
Lume for macOS and Linux VMs on Apple Silicon using Apple's Virtualization.Framework, and Cua-Bench for evaluating computer-use agents on OSWorld, ScreenSpot and Windows Arena

Requirements

The driver install script — `curl -fsSL https://cua.ai/driver/install.sh` on macOS and Linux, or the PowerShell equivalent on Windows — followed by the post-install steps. The sandbox SDK is a separate `pip install cua` and needs Python 3.11 or later. MIT licensed; some optional components carry their own licences.

Setup effort

Build from source — clone the repository and build it, then point your client at the binary