CUAWright uses coding models to build browser automation with inspectable, reusable Playwright scripts. It suits developers working on automation, QA and data entry, with code and logs they can review and maintain. Setup and model-provider costs depend on the chosen environment.
Desktop execution uses OSWorld-V2 in an Ubuntu Docker VM. It should not be described as direct automation of arbitrary Windows or macOS desktops. The browser runtime is MIT-licensed; the imported desktop runtime uses Apache-2.0.
Is CUAWright a separate product from Webwright?
It is the renamed project, with browser and desktop entry points. Existing Webwright commands and Python imports remain compatible.
Does every desktop run receive a score?
Custom desktop tasks have no evaluator or score. The official benchmark workflow requires gated task assets and runs an evaluator.
Who is it for?
Developers building inspectable automation and researchers evaluating computer-use agents. It requires setup and is less suitable for someone seeking a ready-to-use consumer assistant.
The official repository, NOTICE and desktop documentation were checked on 13 October 2026. AICoolies has not independently tested these workflows or reproduced the reported benchmarks.
