winhands

Computer use for Windows: a small MCP server that lets Claude (or any MCP client) operate any desktop app, draw in canvases and drive games.

Install

Windows 10 2004+ or 11, and Claude Code or any MCP client. In PowerShell:

winget install astral-sh.uv    # once; installs uv, which brings its own Python (check the id with: winget search uv)
uv tool install winhands
winhands setup

winhands setup registers the MCP server in Claude Code (user scope) and installs the skill into ~/.claude/skills/winhands. If claude is not on your PATH it prints the config snippet for other MCP clients instead. Then restart Claude Code and ask it: "open Notepad and type hello".

Already have Python 3.11+? pip install winhands, then winhands setup.

• Update: uv tool upgrade winhands
• Uninstall: winhands setup --remove, then uv tool uninstall winhands
• See what setup would do without changing anything: winhands setup --dry-run

Alternative: one-line installer

Installs uv if missing, then does the same as the commands above. You can read the script first; it needs no admin rights.

irm https://raw.githubusercontent.com/zredo19/winhands/main/install.ps1 | iex

To uninstall with it:

& ([scriptblock]::Create((irm https://raw.githubusercontent.com/zredo19/winhands/main/install.ps1))) -Uninstall

Prefer to do it by hand from the repo? See Manual install.

The overlay while Claude controls the PC: edge glow and status banner in four states (acting, thinking, paused, stopped).

The agent cursor at 2x: idle, moving (trail) and click (ripple), in the Claude, Antigravity and Codex palettes.

Its design follows the computer-use harness of OpenAI's GPT-6 Astra in the Codex/ChatGPT desktop app, and goes past it where that harness is weak (no key holds, no raw mouse, no local control loops):

1. Accessibility tree first, pixels when needed. A Codex-style UI Automation tree (12 button "Save" (focused)), plus screenshots automatically when a window is canvas-like (games, Paint), OCR, grid rulers, Set-of-Marks and burst montages.
2. Code mode. The model writes Python in a persistent REPL and batches many actions per turn, including local perceive→act loops that run at 10–30 Hz without model round-trips.
3. Built-in validation. Every run returns a UI diff (or the full tree of a new dialog).
4. Fail closed. Element ids are bound to the latest snapshot; changed or vanished targets raise StaleTarget instead of clicking the wrong thing.
5. Takeover UX and safety. An animated edge glow and status banner (acting, thinking, stopped; with the Ctrl+Alt+Q hint) on the controlled monitor, plus an agent cursor with a motion trail and click ripple where real input lands, all in the client's colour and excluded from screenshots. Ctrl+LeftAlt+Q kill switch, and an automatic abort when the user touches mouse or keyboard. Held inputs are released on abort, risky actions need confirmation, and some windows are denylisted.
6. Memory. Per-app notes and reusable skills (Voyager-style) that persist across sessions.

Tools (2, ~815 tokens of schema)

Input uses SendInput with scan codes (DirectInput/raw-input games see real keys) and raw relative mouse for game cameras. Window capture uses PrintWindow(PW_RENDERFULLCONTENT), so covered windows are read correctly. Apps launched with app() survive the MCP session (launched through WMI, outside the client's kill-on-close job).

End-to-end results (real desktop, Windows 10, Spanish UI)

Benchmark vs Windows-MCP

The same tasks run on the same machine, with each server driven the way an optimal agent would use it. There were 3 reps per task. Script: bench/bench.py, raw data: bench/results.json.

The tool schema is sent on every turn: winhands uses about 815 tokens and Windows-MCP about 4,780.

Notes on method:

• Tokens are estimated: text is chars / 3.5, and images are ceil(w/28)·ceil(h/28) per Anthropic's vision docs.
• Tool time excludes model thinking. Each extra call also costs a model turn in a real agent.
• Windows-MCP's snapshot lists only interactive elements of the focused window, so reading the calculator result needs a screenshot. Its failed rep missed the calculator in that snapshot.
• Windows-MCP dropped rapid consecutive clicks. A 1 s pause between its clicks, not counted, mimics agent pacing.

Manual install

Windows 10 2004+ / 11, Python 3.11+. To run the development version from GitHub with uv:

uv tool install --compile-bytecode git+https://github.com/zredo19/winhands
winhands setup

winhands setup registers the MCP server and installs the skill (perceive → batch → validate, game guidance, safety rules). To do those two steps yourself: claude mcp add winhands --scope user -- winhands, and copy winhands/SKILL.md to ~/.claude/skills/winhands/SKILL.md.

Any MCP client works: the server command is just winhands (stdio). Set WINHANDS_OVERLAY=0 to hide the overlay, WINHANDS_LINGER=45 for how long it stays up between actions and WINHANDS_SHARED=0 for strict mode. From source: pip install -e . in a venv, then point the client at python -m winhands.

Safety

• Ctrl + Left Alt + Q aborts the running action. AltGr+Q still types @.
• Shared mode (default): the user can keep working while a run acts through UIA patterns. Physical input that collides with a run driving the real mouse/keyboard aborts it (UserInterrupt); real input waits for the user to go idle first. The tool never fights the user for control. WINHANDS_SHARED=0 restores strict mode.
• The low-level keyboard/mouse hooks behind both features exist only while a run is acting. Nothing hooks system input while the server idles in a Claude Code session.
• Keys are only sent to a verified foreground target. All held keys and buttons are released on any abort.
• sh() and clicks on elements named like Send, Buy, Pay, Delete, Install or Allow require run(..., confirm=True).
• Denylisted windows (password managers, banking) are refused.
• run executes arbitrary Python: the same privilege as Claude Code's shell tool, gated by the client's permission prompts.
• Online games with anti-cheat: automation may violate their terms and risk a ban. winhands makes no attempt to hide or evade detection.

Limits

• The harness makes a model faster and cheaper on the desktop. It does not make it smarter.
• Reflex gameplay needs local loops written by the model, or pausing the game. Model round-trips take seconds.
• Exclusive-fullscreen games may capture black; use borderless mode. Windows OCR skips isolated single characters and glyphs touching strong borders. Template matching is exact-scale.
• Elevated (admin) windows ignore input from a non-elevated server (Windows UIPI).

Tests

python -m pytest -q