Cua

• Letting Claude Code or Codex build a desktop app and then operate it to QA the result
• Giving an agent its own macOS or Linux Space with a teleported, signed-in Chrome session while you keep using your own Mac
• Running computer-use evals in parallel on warm pools of Linux and Windows machines
• Writing a verifiable desktop task in Cua Bench and scoring several agents or models against it
• Starting a disposable Linux desktop in a GitHub Actions job to run GUI tests with screenshots
• Turning a spare Mac mini into a host that your laptop and agents create Spaces on

Cua is worth a close look if your agent already writes code well but stalls whenever the job needs a GUI. The MIT-licensed driver is the most broadly useful piece: it plugs into the agents people already run and works in the background instead of hijacking the screen. Teams doing computer-use research get sandboxes, fleets, and a benchmark harness from the same vendor. Know which licence covers which piece, though, since Spaces is FSL rather than MIT. Check the platform matrix for the desktops you need, and verify cloud macOS availability directly, because the homepage and the Fleets docs disagree. Treat Teleport and the hosted relay as trust decisions, not conveniences.

About Cua

Cua does not ship its own agent. It supplies the computer and the hands, and you bring the agent and the model. The core piece is Cua Driver, an MIT-licensed binary that reads each window's accessibility tree and screenshot and then clicks, types, scrolls, and drives browser tabs, in the background where the operating system allows it, so the person at the keyboard keeps their cursor and focus while the agent gets its own cursor overlay. Agents reach it as an MCP server, through a CLI, or through Python and TypeScript SDKs, and the docs give setup for Claude Code, Codex, Cursor, Antigravity, OpenClaw, Hermes, OpenCode, Factory Droid, Qwen Code, Kimi Code, and Pi; the harness owns the model and the loop, and Cua Driver never calls a model API. When an agent should not touch your own machine, the Cua SDK and `cua` CLI create disposable Linux, Windows, macOS, and Android environments locally (Docker, QEMU, Lume, or an Android emulator), on any machine running Cua's guest daemon, in your own AWS, Google Cloud, or Modal account, or in Cua's cloud, where Cua Fleets adds named pools, warm replicas, claims, and Terraform for evals, reinforcement learning loops, and batch rollouts, billed per vCPU-hour and GiB-hour. Cua Spaces, released at the start of October 2026, is a free Mac app that lists those desktops in the menu bar, streams them live, gives you and each agent a separate cursor on the same desktop so you can step in and hand it back, and can teleport a signed-in app session such as Chrome or Slack into a Space after you approve each item. Lume manages macOS and Linux VMs on Apple silicon, and Cua Bench runs verifiable desktop tasks and adapters for benchmarks such as OSWorld-Verified and WebVoyager. The tradeoffs: this is developer infrastructure rather than a finished assistant, so you need an agent harness and model credentials of your own; Cua Spaces is source-available under FSL-1.1-MIT rather than MIT and needs macOS 26 for the Mac app; driver coverage on Linux depends on the toolkit and compositor; Cua's cloud runs amd64 images, and its docs say macOS is not offered on the cloud runtime even though the homepage advertises macOS and Android fleets; and the Spaces relay terminates TLS, so it can read streams that pass through it.