The app has no API, so the agent uses the desktop's

The accounting tool your back office depends on is a fifteen-year-old Linux GUI app. No API, no CLI, no export flag. A human clicks through it every Friday.

Now an agent does. It finds the window with POST /api/v1/display/window/search, takes GET /screenshot?base64=true, and hands the image to a vision model to locate the export dialog. One POST /api/v1/display/input/batch call clicks, types the filename, presses Enter: with a closing screenshot to verify. If a run dies mid-click, POST /input/reset releases held keys before the retry.

No automation framework was installed. The desktop ships with its own REST surface, and every application inherits it.