cua.

Computers for computer-use agents

You bring the agent.
Cua brings the computer.

Desktop automation, isolated cloud desktops, local macOS VMs, and a benchmark harness. Four pieces you can take separately.

The division is the product. The project says it plainly: you bring the agent and model, Cua provides the computer and the automation tools.

fleet / desktop pool

Nothing claimed
linux
idle
linux
idle
macos
idle
windows
idle

Illustrative pool. Claiming one leaves the rest untouched, which is how the SDK describes it.

Driver
Operate native apps on macOS, Windows and Linux, plus browsers.
Fleets
Isolated cloud desktops claimed from a pool, run and screenshotted.
Lume
Local VMs for macOS and Linux on Apple Silicon.
Bench
Tasks and trajectories you write, verify and export.

Pools can retain paid capacity after a claim ends, so the tutorials include cleanup steps.
Local sandboxes and cloud Fleets share one SDK, but credentials, images and runtime requirements differ. MIT licensed.

Why a bench ships with the drivers

Tasks you define

A simulated task is written and verified before any agent is scored against it.

Trajectories you keep

Runs export, so an evaluation can be re-read rather than remembered.

Data, not just a score

The project frames the same harness as a source for training and evaluation.