Summary
Reads the UI tree instead of guessing from a screenshot.
Click targets come from Windows UI Automation, so a tap lands on the control rather than near it, and most steps return text instead of an image. The trade is platform: this runs on Windows x64 only.
What it is
A single-binary MCP server for automated desktop UI testing on Windows. It gives an agent the screen, the window list, the Windows UI Automation tree and OCR, plus mouse and keyboard control.
What you get
- Full-screen and single-window screenshots (take_screenshot, take_window_screenshot)
- The actual UI structure, with roles, states and coordinates read through Windows APIs rather than estimated from a picture (list_ui_elements)
- Window enumeration and focus (list_windows, focus_window)
- Mouse click, move, drag and scroll at exact coordinates (mouse_click, mouse_move, mouse_drag, mouse_scroll)
- Typed text, single key presses and shortcut combinations (keyboard_type, keyboard_press, keyboard_shortcut)
- OCR with word-level coordinates, over the whole screen or one region (read_screen_text, read_region_text)
Requirements
Windows x64; macOS and Linux are listed as planned. One binary, with no runtime and no dependencies to install.