trycua/cua

The problem

Building a computer-use agent means solving a large amount of infrastructure before you get to the agent itself: a machine to run on, a way to open apps and read the screen, a way to deliver input, and a way to reset between runs. Most teams end up with a fragile patchwork of VM scripts and screenshot loops that only works on their own laptop and doesn't survive parallel evaluation. Cua packages that layer as a set of components: a desktop app that hosts agent "Spaces", a driver that inspects and operates native apps, a VM manager for Apple Silicon, an SDK/CLI for sandboxes, plus a small decision model and an evaluation harness.

What's different

The distinguishing move is that the SDK has no agent layer inside. Sandboxes are exposed as ordinary runtimes you drive from your own harness, so you keep whatever planning loop you already trust instead of adopting a framework. Second, Cua Driver targets native desktop apps on all three platforms and can deliver input in the background without stealing the user's pointer, which is unusual for GUI automation libraries. Third, Lume builds macOS VMs on Apple's Virtualization.Framework, meaning macOS itself — a platform most agent sandbox stacks simply cannot offer — is a first-class target. That combination is broader and lower-level than browser-only automation such as Playwright, and more platform-native than generic container sandboxes.

Who it's for

Teams building or evaluating computer-use agents who need macOS environments, parallel sandboxes, or reproducible desktop sessions, and who already own their agent logic. It fits researchers generating trajectories and platform engineers standing up an internal sandbox fleet. Look elsewhere if you only need browser automation on Linux — Playwright is lighter — or if you want an end-to-end hosted agent product with no infrastructure work. The README is long and marketing-shaped in places; details on multi-host fleets and Windows Spaces are thinner than the macOS story.

Cost of adoption

Spaces requires macOS 26 or later on the host, with signed installers and permissions to grant. Lume VMs are Apple Silicon only. The CLI and SDK span Python, TypeScript and Rust; Swift and Kotlin bindings are explicitly unreleased. A single-host setup can be done in a couple of commands, but a real fleet means provisioning machines, a relay, and network configuration. Licensing is split: the repository is MIT, while the Cua Spaces app is source-available under FSL-1.1-MIT, which is not the same as MIT. Spaces is free for individuals, with paid Pro and Teams tiers, so commercial use needs a licensing review.

Verdict

This is a young project — created in early 2025 — that has already pulled in roughly 30.8k stars and 2.2k forks and added about 4.9k stars in the last month, landing at #5 on the monthly Rust board. That is fast growth for a stack this broad, and the fact that it never appeared on the daily trending board suggests a steady build rather than a one-day spike. The architecture — small composable binaries with no opinion about your agent — is the right shape for this problem, and the macOS VM capability is something few competitors offer. The risks are maturity and cohesion: several components are still at sub-1.0 versions, and the mixed MIT/FSL licensing plus paid tiers means you must check the commercial terms before building a product on Spaces. Worth a serious evaluation if computer-use agents are core to your roadmap; not worth it if desktop automation is a side feature.