Your Team's Computers Are Already Powerful Enough to Host Agent Desktops
You've probably got a few machines sitting around—a Mac mini under someone's desk, a spare laptop that only gets used for testing, maybe a build server that idles most of the day. Meanwhile, you're trying to figure out where to run agents that need a real desktop to click around in. Cua is a project that asks a simple question: why not use the computers you already have?
What It Does
Cua is a collection of tools for building and running computer-use agents—agents that operate graphical desktops rather than just calling APIs. The headline piece is Cua Spaces, a desktop app that "combines your existing computers into one supercomputer." In practice, that means each desktop your agent uses is a Space: a macOS VM built locally on your Mac, shared by your team and its agents.
The repository bundles several related components. Cua Driver handles desktop automation—inspecting and operating apps on macOS, Windows, and Linux. Lume spins up local macOS and Linux VMs on Apple Silicon. CUA-S1 provides small, specialized models for computer-use decisions. And Cua Bench lets you create tasks, evaluate agents, and export trajectories so you can actually measure whether any of this works. There's also a Cua SDK and CLI for installing cua and creating local sandboxes.
The project is explicit about the division of labor: you bring your own agent and model (or try CUA-S1), and Cua provides the computer and the automation tools.
Why It's Cool
The pitch isn't a new model or a new agent framework—it's infrastructure. That's what makes it worth a look.
-
It meets agents where they already work. The README frames this as "Computer-Use 2.0": an agent moving between code, APIs, and graphical interfaces within the same task. That's a realistic picture of how these agents actually operate. Sometimes the fastest path is an API call; sometimes it's clicking through a UI that has no API at all. Cua tries to support both without forcing a choice.
-
The hardware story is genuinely practical. Agent workloads are bursty and often need real OS environments. Renting cloud VMs for that gets expensive fast. Combining machines you already own into a shared pool for your team and its agents sidesteps a lot of that cost and procurement friction.
-
The pieces are separable. You don't have to adopt the whole stack. If you just need to automate a desktop app, Cua Driver stands alone. If you want local VMs on Apple Silicon, that's Lume. If you're evaluating agents, there's Cua Bench. The monorepo structure means you can pick the tool that solves your problem today.
-
Evaluation is a first-class concern. A lot of computer-use projects hand-wave at benchmarking. Cua Bench explicitly covers task creation, agent evaluation, and trajectory export. If you're serious about knowing whether your agent is good, having that in the same repo as the runtime is a sensible design decision.
How to Try It
The README lays out a path per component, each with its own quickstart. Pick the one that matches what you're doing:
- Cua Spaces: Install the app and run your first Space. Follow the quickstart.
- Cua Driver: Operate Calculator and verify its result. Follow the quickstart.
- Lume: Create a Tahoe VM and connect over SSH. Follow the quickstart.
- Cua SDK and CLI: Install
cuaand create a local sandbox. - Cua Bench: Create and verify a simulated task. Follow the quickstart.
You can grab Cua Spaces from spaces.cua.ai, and the full source lives at github.com/trycua/cua. Documentation is at cua.ai/docs, and there's a Discord if you get stuck.
Final Thoughts
Cua is best suited to teams already experimenting with computer-use agents who've hit the wall of "where do these things actually run?" If you've got spare Macs and a need for sandboxed desktops, the Spaces concept is a pragmatic answer. The collection of separate tools—Driver, Lume, Bench—also means you can start small rather than committing to an entire platform. It's infrastructure work, not magic, but infrastructure is often exactly what's missing.