omarch-eye focuses the Omarchy window you look at
A new beta plugin by OmaCheese uses a webcam above your screen to work out which Hyprland window you are looking at and gives it focus, with everything processed locally.
Omarchy already lets you switch windows without the mouse. omarch-eye, a new third-party plugin by OmaCheese, goes one step further: “Look at a window. It has focus.”
A webcam above the monitor watches your eyes, the software estimates where you are looking, and the matching Hyprland window gets focus. Typing pauses it, the mouse always wins, and if it picks the wrong window, a quick glance up at the camera and back moves focus to the next likeliest one. It also learns how you sit from your corrections and typing.
Under the hood it uses OpenCV and MediaPipe’s face and iris landmarks, plus two models that read the eyes at full resolution. A 15-dot calibration, with an optional refine step that follows the mouse pointer, picks whichever of five models predicts best for you. On the author’s desk, the best model had an error of 4.6% of the screen width.
The README is upfront about limits. It chooses between windows, not buttons, works best with a few large windows, is weaker on windows stacked top and bottom, and supports one monitor. It uses about a third of a CPU core for the face model, and it has only been used live on one desk, with a phone streamed through Flux as the camera.
On privacy, the author says omarch-eye’s own code makes no network requests once installed, camera frames stay in memory and are never saved, and calibration keeps eye descriptors rather than pictures.
It needs Omarchy 4 and a camera above the monitor. Install it with omarchy plugin add https://github.com/OmaCheese/omarcheye.git --enable, then click the eye icon in the bar to run the setup script, which installs packages and a Python environment of about 700 MB. Left-click the eye to turn tracking on or off; right-click opens a menu with calibration, test, preview and camera view.
The plugin is at version 0.8.0 and described as beta. It is MIT-licensed and community-made.
