Cua Rebuilds Computer-Use Agents Around the Window, Not the Desktop

Cua, an open-source project for letting AI agents operate a computer the way a person would, added 609 GitHub stars in the 24 hours before this article was reported, part of a run that has taken the repository to 25,702 stars and 1,768 forks since it was created on Jan. 31, 2025. The project is MIT-licensed and describes itself on GitHub as a way to "scale computer-use 2.0 with open-source drivers, cross-OS fleets, and benchmarks for training, evaluation, and data generation." Stars measure attention, not whether the software works as advertised, so the more useful question is what "computer-use 2.0" actually changes and where the project's own numbers say it still falls short.

What is "computer-use 2.0"?

Computer-use is the general term for letting an AI agent operate a desktop by reading the screen and issuing clicks and keystrokes, rather than calling an API built for it. Cua's engineers, Francesco Bonacci, Dillon DuPont and Robert Wendt, laid out the "2.0" framing in a July 13, 2026 post written up after a talk at AI Engineer World's Fair: "computer-use is becoming a tool inside coding agents like Claude Code and Codex, and general agents like Hermes, OpenClaw, and Pi." The distinction from earlier computer-use tooling is scope. Instead of an agent taking a screenshot of an entire desktop and guessing at pixel coordinates, Cua's approach gives the agent window-level commands, the same view and controls a person gets when they click into one application rather than staring at the whole screen.

The Driver: one interface across three operating systems

The component doing that work is Cua Driver, which the project's README says can "inspect and operate native desktop apps and browsers on macOS, Windows, and Linux" through a single Rust interface. On macOS specifically, the July blog post says the Driver uses Apple's undocumented SkyLight framework to handle background input, and supports a synthetic multi-cursor with session IDs tied to recordings and actions so multiple agents can act on the same machine without colliding. Cua says it validates the Driver with a regression matrix of more than 500 behavioral checks across platforms and application toolkits. The README lists Claude Code, Codex, Cursor and OpenClaw as agents already wired to it; the July post adds Hermes, Pi, Clicky, Qwen Code, H Company and Factory's Droid to that list.

Component	Platforms	What it does
Cua Driver	macOS, Windows, Linux	Window-level control through one Rust interface
Cua Fleets	Linux, cloud-hosted	Isolated cloud desktops agents can drive remotely
CUA-S1	Model, not an OS	Small models for fast, bounded computer-use decisions
Lume	macOS, Linux on Apple Silicon	Local VM management via Apple's Virtualization Framework

A vendor benchmark that says the gap is still wide

Cua's July post includes a benchmark it ran itself on tasks built around KiCad, electronic design software, and the results should be read as vendor-supplied rather than independently verified. On those tasks, GPT-5.5 completed 6 of 25, and Gemini 3.5 Flash reached 5 full solves with 3 partial completions. No model tested crossed even one-third completion, and every agent failed outright on tasks that started from a blank canvas rather than a partially built file.

1. GPT-5.56 tasks
2. Gemini 3.5 Flash5 tasks

Source: Cua, 'Computer-Use 2.0,' cua.ai/blog, July 13, 2026

Where the cross-OS pitch is thinnest: its own home platform

The most-discussed open issue on the repository, with 13 comments, is a complaint that the published Docker image trycua/cua-ubuntu:latest has no build compatible with arm64, the architecture behind every Apple Silicon Mac. That lands squarely on Cua's own strongest claim, since Lume, its tool for running local macOS and Linux virtual machines, exists specifically to target Apple Silicon through Apple's own virtualization framework. A developer running Cua on an M-series Mac and reaching for the containerized Ubuntu environment hits an architecture mismatch the project's marketing does not mention. Licensing carries a smaller wrinkle in the same direction: Cua's own code is MIT, but the README lists an optional dependency, ultralytics, under AGPL-3.0, a license that can force anyone who modifies and redistributes code using it to release their own source under the same terms, a very different obligation from the permissive license on the rest of the project.

None of that makes the underlying idea wrong. Agent frameworks racing toward window-level, cross-OS control are proliferating fast enough that another open-source AI-agent tool gained more than 2,000 GitHub stars in a single day earlier this month, and orchestration is being rebuilt at the infrastructure layer too, the way Google rebuilt its own agent orchestrator around Redis instead of Kubernetes. What Cua's issue tracker shows is that a project can win the framing argument, that computer-use should operate at the window level, while still not having shipped the cross-OS parity its pitch promises, 1,036 open issues into a project that is a year and eight months old.

Sources

1. trycua/cua, GitHub
2. Computer-Use 2.0, Cua