HN sees Cua as the missing layer for background computer-use on macOS

The Hacker News reaction to Cua was less “cool demo” than “someone finally built the missing plumbing.” In the Show HN thread, Francesco Bonacci framed the problem in blunt terms: most GUI automation still hijacks the human session. The cursor jumps, focus changes, windows come forward, and the user has to stop working while the agent clicks around. For HN readers already watching coding agents and computer-use systems spread, that limitation is no longer cosmetic. It is the thing that stops one machine from being shared safely between a person and multiple agents.

The project, cua-driver, is an open-source macOS driver meant to let an agent click, type, scroll, and inspect a real Mac app while the user stays in the foreground. The accompanying technical writeup explains why this is harder than it sounds. CGEventPost goes through the HID stream and warps the cursor. CGEvent.postToPid avoids that but Chromium can reject those clicks as untrusted. Raising the target window first makes the click work, but destroys the whole point by stealing focus and sometimes dragging the user across Spaces.

Cua’s answer is to lean on undocumented macOS internals. The writeup centers on SkyLight’s SLEventPostToPid, a focus-without-raise pattern borrowed from yabai, and a private accessibility hook that keeps Electron accessibility trees alive even when a window is occluded. There is even a “primer click” trick: one synthetic click at (-1, -1) to move Chromium’s user-activation gate, then the real click lands without the window ever coming forward. Bonacci argues that background computer-use should be commodity infrastructure for any harness, not a proprietary capability locked inside one vendor’s agent product.

HN commenters largely bought that framing. One ex-Apple engineer said the implementation looked strong and called out the big practical win: running multiple UI automation flows at the same time. Others immediately pushed on what comes next, from telemetry defaults to whether Windows or Linux will need similar primitives if agent workflows keep growing. Community discussion noted another important detail from the writeup too: the current approach still has limits, including Chromium right-click issues and canvas-heavy apps that sometimes need frontmost activation. Even with those caveats, the thread treated Cua as more than a neat Mac trick. It looked like early infrastructure for a world where “computer use” stops meaning one remote demo and starts meaning concurrent, everyday agent work on a real desktop.