Show HN: I spent 3 months making desktop automation stop lying to AI agents
TL;DR
A developer spent three months building agent-desktop, a command line automation framework for desktop applications. It works like Playwright but targets native and Chromium based desktop apps instead of the browser. The core problem it addresses: accessibility trees often misreport the real UI state, which makes AI agents fail on long horizon tasks. macOS support is generally available, with Windows and Linux in progress.
Nauti's Take
The opportunity is real: desktop automation lags far behind browser agents, and a Playwright style CLI for native apps fills an obvious gap. The risk sits in the claim itself, because accessibility trees break differently in every app and a Show HN demo is no substitute for weeks of production use.
Teams automating internal macOS tools can start testing now, while Windows and Linux shops should wait for the promised releases.
Sources