An agent drives a Windows application through UI Automation, every action is written to a journal with exact timings, the window is captured with Windows Graphics Capture, and the video is assembled from that journal rather than from the footage — so the camera starts moving 0.55 s before each click.
What is inside
- 19 core modules (5588 lines), 6 test scripts (1093 lines), a 26-file sound bank
- 20 MCP tools for driving and recording the desktop
- Camera, pause-collapsing timeline, compositor, sound design, optional narration
Getting started
/plugin marketplace add <path-to-clone>
/plugin install screencast-desktop
/screencast-desktop:setup
The UI Automation binary is not vendored — server/fetch_ui_binary.py downloads it from sbroenne/mcp-windows (MIT) and verifies the checksum.
Known limitations are listed honestly in the README, and what is planned about them — with the measurements behind each item — is in ROADMAP.md. Notably: no drag primitive yet, keyboard events do not reach the journal, and the frame origin is off by the title-bar height on windows that have one.