Most computer-use agent demos break the moment they touch a real desktop: the model grabs your mouse, fights the window manager, and can't be reproduced. Cua's bet is that the hard part isn't the agent's reasoning — it's the environment, so it ships the plumbing (sandboxed OS instances, a uniform control API, and benchmarks) and lets you bring whatever model does the thinking.
What Sets It Apart
- One SDK, many backends: the same code drives a local QEMU VM, an Apple-Virtualization macOS guest, or a cloud desktop on cua.ai — so a laptop prototype and a production fleet share one interface.
- Background drivers run automation on native macOS/Windows/Linux apps without stealing focus or the cursor, so a human and the agent can share a machine instead of you watching a hijacked screen.
- Near-native speed: containers report up to 97% of native CPU on Apple Silicon, making per-step screenshot-and-act loops cheap enough to actually train and evaluate on.
- Cua-Bench bundles RL environments and a scoring harness, turning "did the new model get better at booking a flight?" into a measurable question rather than a vibe.
Who It's For
Great fit if you're building or researching agents that click, type, and navigate real operating systems and you want reproducible environments plus an MCP bridge into Claude Code or Cursor. Look elsewhere if you just need headless browser scraping (Playwright is lighter) or a turnkey consumer assistant — Cua is infrastructure you assemble an agent on top of, and the Linux backend is still pre-release.
