WinCtl

[WinCtl]

WinCtl

Full Windows desktop access for Claude and other MCP clients
See the screen, read and drive real application UIs, manage windows and processes, work with files, and run shell commands — all behind a tiered permission model with audit logging.

[npm] [license]

Why this connector

Most desktop automation servers hand the model a screenshot and a click(x, y) tool, which breaks the moment a window moves. This one is built around the things that actually make Windows automation reliable:

• UI Automation first. uia_snapshot reads an application's accessibility tree, so Claude can act on "the Save button" rather than on coordinates that go stale. uia_invoke and uia_set_value drive controls directly.
• Correct coordinates, always. The process declares per-monitor DPI awareness before touching any UI API, and reports the DWM frame bounds rather than the raw window rect (which includes an invisible ~7px resize border). Without this, every coordinate on a scaled display is silently wrong — on a 200% display, GetWindowRect reports 1351px where the true edge is 2641px.
• Input that actually arrives. Synthetic typing is paced, because batching Unicode key events makes applications drop and repeat characters. Text over 200 characters is pasted via the clipboard instead, and your clipboard is restored afterwards.
• Handles that cannot betray you. Window ids are opaque and fingerprinted; Windows recycles window handles, and a stale one silently retargeting to a different app is exactly how automation closes the wrong window.
• Honest failures. window_focus verifies the foreground actually changed and says so when Windows refuses. Errors explain what to do next rather than surfacing an HRESULT.

Requirements

• Windows 10 (1809+) or Windows 11
• Node.js 20 or newer
• No compiler or build tools — every native dependency ships prebuilt binaries

Install

Claude Desktop (recommended)

Download winctl.mcpb from the latest release and double-click it. Claude Desktop installs it and exposes the permission profile, allowed folders and confirmation settings in its UI.

Claude Code

Install globally first, then register the command:

npm install -g @sitharaj88/winctlclaude mcp add winctl -- winctl

Add --scope user to the second command to make it available in every project rather than just the current one.

Don't use npx -y @sitharaj88/winctl here. It works, but WinCtl depends on prebuilt native binaries (sharp, koffi), and npx re-resolves them on every launch — around 18 seconds versus 3 for a global install. MCP clients give up long before that and report Connection closed.

Any MCP client (manual)

{  "mcpServers": {    "winctl": {      "command": "winctl",      "env": {        "WINCTL_PROFILE": "standard"      }    }  }}

If winctl isn't on your PATH, point at the entry point directly — on Windows a global npm install lands in %APPDATA%\npm\node_modules:

{  "mcpServers": {    "winctl": {      "command": "node",      "args": ["C:\\Users\\<you>\\AppData\\Roaming\\npm\\node_modules\\@sitharaj88\\winctl\\dist\\index.js"]    }  }}

Claude Desktop's config lives at %APPDATA%\Claude\claude_desktop_config.json.

Permission tiers

Every tool belongs to exactly one tier. Tools in a disabled tier are never registered, so the model cannot see them, attempt them, or spend context reading their descriptions.

Profiles bundle these:

# Pick a profileWINCTL_PROFILE=readonly
# …or choose tiers explicitlyWINCTL_TIERS=observe,interact

Configuration

An unrecognised profile name fails closed to readonly with a warning on stderr, rather than guessing what you meant and possibly granting write access.

Tools

40 tools total. Every one carries title, readOnlyHint and destructiveHint annotations.

Example

"Open Notepad, write my meeting notes into it and save to Documents."

Claude will typically:

1. process_start → launch Notepad
2. window_list → find the window and its owning process id
3. window_focus → make sure keystrokes land there
4. input_type → paste the notes via the clipboard
5. input_press_keys ctrl+s, then uia_set_value on the filename field
6. screen_capture_window → confirm the result visually

Safety

• Tier gating is the primary boundary — disabled tools are never exposed.
• Path containment resolves symlinks before checking, so a link inside an allowed folder cannot reach a denied one. System32, WinSxS, Boot and the Startup folder are refused in every profile, including full.
• Confirmation via MCP elicitation before destructive actions.
• Audit log of every call in JSONL, with sensitive-looking arguments redacted. See PRIVACY.md.
• Modifier release on shutdown, so a crash mid-chord cannot leave Ctrl latched down.
• Refuses to terminate its own process or pids 0–4.

Screenshots capture whatever is on screen and send it to your AI provider. Read PRIVACY.md before enabling screen capture on a machine that handles confidential material.

Elevation

Windows blocks a normal-privilege process from automating, capturing or sending input to windows owned by elevated processes (UIPI). To drive an administrator application, run the connector elevated — otherwise those calls fail with a clear explanation rather than silently doing nothing.

Privacy Policy

WinCtl runs entirely on your machine. It has no backend, no telemetry and no analytics, and it makes no network calls of its own. The author receives nothing about you or your computer.

• What it accesses: screen contents, window and application state, system and process information, files within the folders you permit, and the clipboard — each only when a tool is called. It generates synthetic input but never records your real keyboard or mouse activity.
• Where data goes: to the AI client you connect it to, and nowhere else. Anything a tool returns becomes part of that conversation and is therefore subject to that client's privacy policy. A screenshot sends whatever is on screen to your AI provider — consider that before enabling screen capture on a machine handling confidential material.
• What is stored: only a local audit log at %LOCALAPPDATA%\winctl\audit.jsonl — one line per tool call, with sensitive-looking arguments redacted. It is never uploaded. Disable it with WINCTL_AUDIT_DISABLED=1, or delete the file at any time.
• Retention and sharing: nothing is retained off-machine and nothing is shared with third parties.
• Contact: Sitharaj Seenivasan — [email protected]

Full policy: PRIVACY.md

Development

git clone https://github.com/sitharaj88/winctlcd winctlnpm installnpm run build
npm run smoke                 # self-test against the live desktopnode scripts/smoke.mjs --full # full MCP protocol suite, including UI Automationnpm run inspect               # MCP Inspector UI

scripts/verify-interactive.mjs drives Notepad end-to-end — it launches the app, types, verifies the text through UI Automation and cleans up. It uses the real mouse and keyboard, so don't run it while you're using the machine.

Architecture

src/  index.ts              stdio entry point (DPI awareness runs first)  http.ts               streamable HTTP entry point, loopback-only by default  config.ts             profiles, tiers, limits  security/             tier definitions, path containment, audit log, errors  native/    ffi.ts              koffi bindings to user32/kernel32/dwmapi    dpi.ts              per-monitor DPI awareness    handles.ts          opaque, fingerprinted window ids    windows.ts          enumeration, focus, move, state    input.ts            SendInput mouse and keyboard    screen.ts           capture and token-aware encoding    powershell.ts       per-call PowerShell execution    uia.ts              UI Automation bridge  server/    registry.ts         tier gate + confirmation + audit in one place    createServer.ts     assembles tools, resources and prompts  tools/                the 40 tool definitions

Two implementation notes worth knowing if you extend this:

• PowerShell runs one process per call. A long-lived host sounds better, but -Command - buffers all of stdin until the stream closes rather than executing statement by statement, so it never returns a result. Per-call processes also mean a hung script can't wedge later calls. Commands are passed as -EncodedCommand (base64 UTF-16LE), which removes shell quoting as a source of injection bugs.
• The INPUT struct size is asserted at startup. A layout mismatch wouldn't throw — it would send mouse events to garbage coordinates.

Publishing

npm run build && npm publish --access public   # npmnpx @anthropic-ai/mcpb pack                    # build the .mcpb bundlemcp-publisher login github && mcp-publisher publish   # MCP registry

For the Claude Connectors Directory, submit desktop extensions via the extension form.

Author

Sitharaj Seenivasan

• Website — sitharaj.in
• GitHub — @sitharaj88
• LinkedIn — in/sitharaj08
• Email — [email protected]

Issues and feature requests are welcome at github.com/sitharaj88/winctl/issues.

Licence

MIT © Sitharaj Seenivasan