What it does
OSWright gives an LLM control of desktop applications through 43 MCP tools covering screen capture, OCR, Windows UI Automation element lookup, image matching, mouse and keyboard input, window management, clipboard access and app launching. Between actions it keeps a model of the screen, rescans only the regions that changed, and can return what changed instead of a full screenshot. Element lookups use the cheapest method that works.
It suits agents that need to operate desktop software where repeated full-screen screenshots would use many tokens. The publisher reports 7.4x less context than returning a screenshot after every action on an eight-field invoice-to-expense-form task, and a median screen change of 0.012% of pixels on a live desktop. These figures are the publisher's own measurements, and the repository includes benchmark scripts to re-run them.
Running it requires Python 3.10 or newer and an MCP client such as VS Code, Cursor or Claude Desktop, configured with the standard setup or installed with pip. On Windows it uses the built-in OCR engine; on Linux and macOS it falls back to EasyOCR. The repository includes an agent skill, SKILL.md, which sets the delta observation mode, and a test suite of 259 tests in which the desktop-driving tests skip when no display is available.
Limitations and requirements
Requires Python 3.10 or newer and an MCP client. Linux support uses pynput with X11 and macOS support uses pynput with Quartz; Linux and macOS OCR uses EasyOCR as a fallback. Desktop-driving tests require a display. Without the delta observation mode, which the bundled skill sets, every action returns a full screenshot.
Safety
Our team has not reviewed this resource. Only automated rules checked it; resources with high-risk signals are not published. These are signals, not a guarantee that installing it is safe. Try it in a sandbox project first.
- Runs code or changes an assistant's behaviour
Sources
- Official MCP Registry first seen 13 Nov 2026 · last seen 13 Nov 2026 · its page on the source
io.github.Ask-812/oswright
Windows desktop automation that re-reads only what changed: 7x fewer tokens per task.
v0.8.2As shown on registry.modelcontextprotocol.io when we last read it