What is Cua?
Cua is an open-source (MIT license) framework and hosted infrastructure platform that lets AI agents control desktop operating systems, including Linux, Windows, macOS, and Android, inside lightweight virtual sandboxes. It has four main components: Cua Driver handles native desktop interaction via clicks, typing, and accessibility-tree inspection; Cua Sandbox provisions local or cloud VMs through Docker, QEMU, or Apple's Virtualization Framework; Cua Fleets is elastic hosted infrastructure with warm machine pools; and Cua Bench handles cross-surface agent evaluation. The stack targets computer-use agent training, evaluation, and data-generation workloads.
Our verdict
The MIT-licensed core and 20.6K-star GitHub repo make self-hosting straightforward for teams building computer-use agents that need OS-level control rather than just browser control. Paid tiers cover hosted fleets, BYOC, on-prem deployment, and SOC 2 Type I compliance, but they sit behind a request-access flow. The team has three people and was founded in 2025, so its enterprise track record is still limited.
Categories:
Agentic automation tools give an LLM control of a browser to navigate, click, and fill forms toward a goal, rather than following a fixed script. They are newer and less predictable than selector-based scraping, but they handle sites and workflows that are hard to hardcode.
