trycua/cua is an open-source stack that brings drivers, sandboxes and benchmarks together for a computer-use environment in which AI agents perform tasks that people used to do on screen. It is implemented in Rust and licensed under MIT. The repository's latest release is cua-spaces-v0.12.1.
This stack covers, by role, what an agent needs in order to work with a computer. Cua Driver operates native apps and browsers, while Cua Spaces provides the full desktop that the agent will use, on Macs and on devices the user already owns. The repository also contains Lume for local VMs, the decision-making model CUA-S1, and the evaluation tool Cua Bench, together with benchmarks for learning, evaluation and data generation. The Cua SDK and CLI support local VMs, containers, and the device on which cua-spacesd runs. It has 30443 GitHub stars.
There are still points that need checking, however. The repository description foregrounds a sandbox that supports multiple operating systems, but the desktop coverage described in the README extends only as far as Macs and user-owned devices. The scale and license of CUA-S1, the evaluation criteria and results of Cua Bench, and the changelog for the cua-spaces-v0.12.1 release are not given. There are no performance measurements either. The only recorded evaluation score is a momentum of 8.7 out of 10, with no novelty, technical or quality scores. Taking into account that the repository was created on January 31, 2025, it is better treated as a starting point for agent-based desktop automation until its scale and performance evidence can be verified directly.