Skip to content

winuse 0.1.0

Choose a tag to compare

View all tags
@straygizmo straygizmo released this 20 Oct 13:57
· 73 commits to main since this release
v0.1.0
752280b

First release of winuse, an MCP server and CLI that lets an AI observe and operate Windows desktop apps.

Highlights

  • MCP server over stdio (winuse --mcp) with 48 tools: observe (screenshot, UI Automation tree, WebView2 DOM via Playwright over CDP), intent (act, find, check), input, element, window, wait, process, file, assert and scenario tools.
  • Scenario replay without AI (winuse run, scenario_run): recorded locators resolve deterministically; --heal re-grounds act steps whose element is gone.
  • Two decision models: TypeSafe Jev (JEV_API_KEY) or Cloudflare Clef on Workers AI (CLOUDFLARE_ACCOUNT_ID + CLOUDFLARE_API_TOKEN), chosen automatically or forced with WINUSE_DECISION_PROVIDER.
  • Visual grounding for windows without UI elements (VM consoles, emulators, VNC): act / find / check work from a screenshot, return pt{n} refs, and replay from a window-relative point (WINUSE_SCREENSHOTS, WINUSE_VISUAL_LOCATOR).
  • winuse doctor reports the decision model, keys, endpoint, probe, UIA, WebView2 and the Playwright driver; winuse --version prints the version.

Install

  1. Install the .NET 10 runtime (or SDK).
  2. Unzip winuse-0.1.0-win-x64.zip into a folder.
  3. Register with Claude Code: claude mcp add winuse -- <folder>\winuse.exe --mcp
  4. Set JEV_API_KEY, or CLOUDFLARE_ACCOUNT_ID and CLOUDFLARE_API_TOKEN, as user environment variables, then run winuse doctor.

Known issues

  • type_text with Japanese text can be garbled under the Japanese IME (intermittent, under investigation).
  • Clef and visual grounding are covered by unit tests with a fake model; they have not yet been exercised against a live Cloudflare account or a real VM window.

Full changelog: CHANGELOG.md

Assets 3
Loading