WinCU / Windows Computer Use / Computer Action

About Windows computer use: it is also called Codex Computer Action, but in Claude we have always called it WinCU. Why write it? It is like PiUse, which I released earlier, but PiUse was not open source and runs on Linux, while this version is fully open source.

PiUse is a great project, the one I put a cluster of 8 RTX 6000 Pro cards into. It is on hold for now. As for its DK12, its C kernel optimizations and its other advanced features, I will open-source them step by step someday, because the code is simply too big.

The reason for writing it is simple: Codex's computer use is too hard to use, and it breaks with every update. My quota kept dropping day after day. Our daily work can't be paying the company and then repairing its tool for it, can it? So we wrote our own, and it even surpasses theirs. With ChatGPT 6, it has even surpassed Computer Use. As for why, try it yourself.

It supports all AI models. Most of the testing was done by Anthropic's advanced models, but ChatGPT 6 and Chinese models with vision can use it too; in fact, they all can, and I have tried them all. Because it connects through MCP, it is more direct and freer. But please don't modify it too freely: it is an open-source MCP that is even more industrial-grade than Computer Use, a complete Windows computer-control system for stable use. There are no binaries and no compiled .pyc files, so you can use it directly and freely.

One more point: compared with Computer Use, we have no flashy front end, because I don't think Windows needs that kind of thing, such as that transparent glass overlay, the blue one. It is also written directly in Python, not in C. I will keep optimizing it as I keep using it, but I have been using it for industrial-grade work recently, and I think it is already smooth and stable enough, with no cache buildup at all. The whole computer runs very smoothly, unlike before with Computer Use, which had massive memory leaks, all kinds of bugs, and OpenAI's updates to Sky. That was simply terrible.

Overview

An MCP server that lets an AI agent operate a Windows PC: screenshots, mouse, keyboard, clipboard, windows and UI Automation accessibility trees. Input is sent with SendInput, windows are captured with PrintWindow, and accessibility trees are read through UI Automation. Python, stdio transport, version 2.0.0.

Related MCP server: windows-computer-use-mcp

Tools

• Two coordinate frames. Without window_id, an action works on the whole desktop. With window_id, the window is brought to the front first, and coordinates are pixels of that window's own screenshot, which is captured correctly even while the window is covered.

• Screenshots are scaled down so the longest edge is at most max_image_edge (2000 px). The pointer is drawn into desktop screenshots, and off-limits windows are blanked. A screenshot whose PNG is larger than max_png_bytes (300,000 bytes) is sent as JPEG (quality 80).

• Actions answer with one line of text, not an image. computer_batch runs up to 30 steps in order within 75 s, stops at the first step that fails, and ends with a screenshot.

• type takes text in any language. A line break at the very end presses Enter; text with line breaks inside, or longer than 400 characters, is pasted through the clipboard.

• Elements are addressed by accessible name or by index in the latest get_window_state tree. click_element clicks the element's centre with the real mouse and reports which element it clicked.

• Time limit. Each call ends within 100 s. A call stuck in an application that stopped responding is abandoned and its worker thread is replaced.

Refusals

• Off-limits windows: terminals and consoles, Windows security and sign-in prompts, password managers, Start and Windows search, the Run dialog, and Claude's own windows, which are also blanked in screenshots. See denied_apps and denied_window_classes.

• The Windows key is never sent.

• Code editors (click_only_apps) allow only looking, left-clicking and scrolling, because their built-in terminals run commands.

• Banner and Esc. While input is being sent, a banner shows at the top of the screen; it never appears in screenshots. Pressing Esc while it shows refuses input for 12 s.

• The user comes first. If the mouse or keyboard was used in the last 0.8 s, input waits up to 4 s and is then refused (yield_to_user).

• One driver at a time. Input is refused while a window listed in busy_markers is visible, and for session_hold_s (20 s) after another Claude Code session last sent input through win-cu.

• No stuck buttons. A held mouse button is released when Esc is pressed, when a call is cancelled, and when Claude Code exits.

Requirements

• Windows 11 (tested on version 24H2, build 26100)

• Python 3.13 (tested with 3.13.1)

• mcp>=1.23,<2, comtypes>=1.4, pillow>=10 (tested with mcp 1.23.1, comtypes 1.4.16, pillow 12.3.0)

Install

pip install -r requirements.txt

Register the server for Claude Code in ~/.claude.json, under the top-level mcpServers (user scope):

"win-cu": {
  "type": "stdio",
  "command": "C:/path/to/python.exe",
  "args": ["-u", "-X", "utf8", "C:/path/to/win-cu/server.py"],
  "timeout": 150000,
  "alwaysLoad": true
}

Replace both paths with your own. Each new Claude Code session starts its own server process.

Skill

skills/win-computer-use/SKILL.md is a Claude Code skill with win-cu's core notes: coordinate frames, batching, tools, refusals, error messages and configuration. The server's instructions refer to it. Its Other section is left for your own rules. To install it, copy the skills/win-computer-use folder to ~/.claude/skills/.

Configuration

Settings are read from config.json next to server.py, or from the file named by the WINCU_CONFIG environment variable. Each key replaces the default from DEFAULT_CONFIG in wincu/core.py; a list replaces the whole default list. The shipped config.json sets max_image_edge, show_banner and yield_to_user.

Logs

• logs/win-cu.log: one line per call, with its arguments and outcome. It rotates at 1 MB and keeps 3 old files. Typed text longer than 24 characters is logged as its length only.

• logs/owner.json: the session that last sent input.

Tests

Run each script directly, for example python tests/test_unit.py.

• test_tools.py, test_scaled.py, test_tiers.py and e2e_session.py move the real mouse and keyboard. Run tests/desk_free.py first and check its exit code.

• The test windows are C# programs. Build them into tests/ with the .NET Framework compiler:

  C:\Windows\Microsoft.NET\Framework64\v4.0.30319\csc.exe /nologo /target:winexe /out:tests\TestForm.exe tests\TestForm.cs
  C:\Windows\Microsoft.NET\Framework64\v4.0.30319\csc.exe /nologo /target:winexe /out:tests\WebForm.exe tests\WebForm.cs

• The web page check in test_tools.py needs a WebView2Loader.dll, which is not part of Windows. Set WINCU_WEBVIEW2_LOADER to the path of one; without it the check is skipped.

• e2e_session.py runs Claude Code from %APPDATA%\npm\node_modules\@anthropic-ai\claude-code\bin\claude.exe and uses quota for one short session.

• tests/calib.mcp.json has placeholder paths; replace them with yours.

• Test screenshots are written to tests/out/.

Layout