A new open-source project called cua is rapidly climbing GitHub Trending, amassing 383 stars in a single day. The project describes itself as a platform to "give AI agents computers they can use" — providing the infrastructure layer for so-called "Computer-Use 2.0," where AI agents move fluidly between code, APIs, and graphical interfaces within the same task.
What cua provides
The platform bundles four components:
- Cua Driver: An automation driver that lets AI agents inspect and operate native desktop apps and browsers on macOS, Windows, and Linux. It connects via CLI, MCP, or typed SDKs, and supports background delivery so agents can work without stealing focus. - Cua Fleets: Isolated cloud desktops provisioned on demand at run.cua.ai. Code claims a desktop from a pool, runs commands, captures screenshots, and cleans up — useful for testing and data generation at scale. - Lume: A local VM manager for macOS and Linux on Apple Silicon, using Apple's Virtualization.Framework. Users can spin up vanilla macOS or Linux VMs and connect over SSH. - Cua Bench: A benchmark suite for building computer-use tasks, evaluating agents, and exporting trajectories for training — designed to work without requiring a VM, Docker, or model API key.
Why it matters
Computer-use — the ability of AI agents to interact with desktop applications through visual interfaces rather than just APIs — is one of the fastest-moving frontiers in AI tooling. Major labs including Anthropic, OpenAI, and Google have all shipped or previewed computer-use capabilities, but the tooling for evaluating, benchmarking, and deploying these agents across operating systems remains fragmented.
cua's approach of combining open-source drivers with cloud fleets and benchmarks in a single package addresses a real gap. The project already integrates with popular AI coding agents including Claude Code, Codex, and Cursor.
The project is MIT-licensed and welcomes contributions.
