winctl

Give your agent the hands and eyes of a Windows user.

winctl is a local MCP server that exposes the whole desktop: see the screen, move the mouse, type on the keyboard, manipulate windows, launch applications, run PowerShell, and send you a summary when it is done. No driver to install, no service to keep running, no account to create — one Python package, one registration line, and the agent sees your desktop.

Everything stays on the machine. Nothing leaves it, except the summaries you explicitly ask for, to the channel you configure yourself.

---

Requirements

• Windows 10 or 11. The project is ctypes on top of the Win32 API: it runs on no other system, and that is not a temporary limitation.
• Python 3.11 or later.
• An open interactive session. Input injection needs a desktop: no Windows service, no container, no SSH without a graphical session.
• An MCP client. The examples use Claude Code, but the server speaks standard MCP over stdio.

Installation

With uv:

uv tool install winctl-mcp

Or from source:

git clone https://github.com/<OWNER>/winctl-mcp
cd winctl-mcp
uv sync

Check that the command answers:

winctl-mcp --help

Registering with Claude Code

claude mcp add winctl --scope user -- winctl-mcp

From source, point at the virtual environment interpreter:

claude mcp add winctl --scope user -- "C:\path\to\winctl-mcp\.venv\Scripts\python.exe" -m winctl.server

--scope user makes the server available in all your sessions, whatever the working directory. Check it:

claude mcp list

The tools then appear under the mcp__winctl__ prefix — for example mcp__winctl__screenshot. The winctl name is the one you gave to claude mcp add: if you register it under another name, the prefix follows.

Then ask the agent: "take a screenshot and tell me what you see".

---

The tools

See the screen

Tool	Purpose
screenshot	Capture one monitor, all monitors, or a specific region
screenshot_window	Capture a given window, after bringing it to the foreground
screen_info	List the monitors, their coordinates, the cursor position

Two options change everything on screenshot:

• grid: true overlays a grid annotated with real screen coordinates. It is the most reliable way to aim a click to the pixel.
• region_x/y/width/height captures a specific area. Far more legible than a downscaled full screen when small text has to be deciphered.

Mouse and keyboard

Tool	Purpose
mouse_move, mouse_click, mouse_drag, mouse_scroll	Pointer control
type_text	Type text, accented characters and emoji included
press_key, hotkey	Single key or combination (["ctrl","c"])
paste_text	Insert through the clipboard, for long texts
clipboard	Read / write the clipboard

Key names accept English as well as French: enter/entree, escape/echap, delete/suppr, up/haut, down/bas, left/gauche, right/droite, shift/maj, space/espace, tab, f1…f24, win.

Windows, applications, system

Tool	Purpose
list_windows	Open windows, with position, size and process
focus_window, window_action, move_window	Foreground, minimize / maximize / close, move
launch_app	Launch by common name, full path or URL
list_installed_apps	Find the exact name of an installed application
run_powershell	The escape hatch for everything else
wait	Pause between two actions

launch_app accepts "spotify", "chrome", an .exe path or a URL. It chains PATH, the App Paths registry key, then the Start menu — which also covers Microsoft Store applications, since those have no directly addressable executable.

Notifications

Tool	Purpose
send_summary	Send a summary, with a screenshot attached on request
notification_status	Report which channels are configured, and which .env was read
telegram_find_chat_id	Telegram setup helper

send_summary exists for long tasks: the agent works, you do something else, and the result lands on your phone. With no configuration, it falls back to a local Windows notification balloon.

---

Configuring notifications (optional)

Copy .env.example to .env and fill in what you need. No variable is mandatory.

Where winctl looks for the .env

It does not guess — that is deliberate. An MCP server is started by its client, with the working directory of the session; and once winctl is installed as a package, its code lives in a site-packages directory where nobody writes their configuration. Looking "next to the code" or "under the current folder" is only right by accident.

The location is therefore declared, in this order of priority:

1. the path passed to notify.load_env(path) by the calling program;
2. the WINCTL_ENV_FILE environment variable;
3. failing that, .env in the current working directory.

The robust form is the second one, set when registering the server:

claude mcp add winctl --scope user \
  --env WINCTL_ENV_FILE=C:\Users\you\.config\winctl\.env -- winctl-mcp

When in doubt, notification_status reports the file actually loaded, or the list of locations tried.

The .env is reread on every action: completing the file does not require restarting the server. A variable already present in the parent process environment (through --env, for instance) always wins over the file.

Telegram — recommended

It is the only channel that can be set up in two minutes: no OAuth, no domain to verify, no prior review.

1. On Telegram, write to @BotFather, send /newbot, follow the steps.
2. Paste the resulting token into TELEGRAM_BOT_TOKEN in the .env file.
3. Send any message to your new bot.
4. Ask the agent: "call telegram_find_chat_id". Copy the returned identifier into TELEGRAM_CHAT_ID.

Email

Fill in SMTP_HOST, SMTP_PORT, SMTP_USER, SMTP_PASSWORD, SMTP_FROM and SMTP_TO. With Gmail, use an app password, not the account password.

---

Safety, and what this project cannot do

This section is long on purpose. A tool that hands control of a desktop to a language model deserves precision about its edges.

What you are granting. An agent equipped with winctl can do to your session anything you could do there: read your open windows, type into your messaging app, delete files through run_powershell. There is no sandbox. That is the point of the tool, and it is also its risk: do not register it on a machine whose contents are not yours.

Observation mode. Set WINCTL_READONLY=1 in the .env to allow only captures and reads. Keyboard, mouse, window moves, application launches and PowerShell are then refused with an explicit message. It is the right setting for a first hands-on, or to let an agent diagnose without acting.

Windows running as administrator. Windows forbids a non-elevated process from sending input to an elevated window: that is the UIPI mechanism, and there is no clean way around it. If an application started as administrator ignores keystrokes, this is why. The MCP client itself must be started as administrator to drive it — with everything that implies.

Windows only. The winapi module is a direct ctypes binding on user32: SendInput, EnumWindows, SetForegroundWindow. Nothing is portable, and nothing is meant to become so.

A graphical session is required. Without an interactive desktop, SendInput has nowhere to write. A server with no open session, a locked session or a secure desktop (the UAC prompt, Ctrl+Alt+Del) are out of reach: those are separate desktops, protected by the system.

Targeting the right window. The typing tools accept a window parameter. The target window is brought to the foreground and the action is cancelled if focus fails, rather than typing into the active application by mistake. Use it systematically for any keystroke that matters.

Multi-monitor. Coordinates are in real pixels of the virtual desktop and may be negative: a monitor placed to the left of the primary one has x coordinates below zero. Call screen_info when in doubt rather than assuming.

Games and exclusive fullscreen applications. They often capture input at a level SendInput does not reach, and their rendering can escape screen capture. Borderless windowed mode generally works.

Long texts. Prefer paste_text over type_text beyond a few hundred characters: it is instantaneous and immune to applications that drop keystrokes.

---

Notable implementation details

These choices deserve an explanation, because they fix real failures observed in use. If you hesitate to trust this code, this is the section to read.

Character-by-character Unicode typing. The keyboard goes through SendInput with KEYEVENTF_UNICODE, which makes typing independent from the keyboard layout — "é", "ç" or "€" come out identical on AZERTY and QWERTY. But modern controls (WinUI / Text Services Framework, including the Windows 11 Notepad) lose the contents of bursts: past the first few characters, they all take the value of the last one. One SendInput call per character, spaced by 3 ms, delivers the text intact everywhere.

Focus verified, never assumed. Windows only grants SetForegroundWindow to the process that emitted the last user input. The server makes itself eligible by synthesising an ALT keystroke, attaches its input queue to that of the active window, then checks the result and retries. Typing into the wrong window is the worst possible failure for this tool: a clean error beats a message sent to the wrong conversation.

DPI awareness enabled at startup. Without SetProcessDpiAwarenessContext, Windows lies about coordinates and screen sizes as soon as one monitor is not at 100% scaling: the agent would aim right and click beside.

Grid in real coordinates, not image pixels. Captures are resized to fit in the model's context. Annotating the grid with image coordinates would force a conversion back, hence a rounding error per click; the labels therefore carry the clickable value directly. The summary attached to every capture restates the conversion formula, for the cases where the model reasons on the image anyway.

Interpolated mouse movement for drag and drop. An instant jump from one point to another is ignored by most applications, which expect a sequence of movements to arm the drag. mouse_drag interpolates the path.

The .env is reread on every action, never cached. Completing the configuration therefore does not require restarting the server — which matters when it runs under a client one is reluctant to restart. Empty values are ignored, and a line removed from the file removes the variable.

---

Environment variable reference

All of them go in the .env (see .env.example). None is mandatory.

Variable	Default	Purpose
WINCTL_ENV_FILE	—	Path of the .env to load. Read from the environment, not from the .env
WINCTL_READONLY	—	1 blocks keyboard, mouse, windows, applications and PowerShell
TELEGRAM_BOT_TOKEN	—	@BotFather token, for send_summary
TELEGRAM_CHAT_ID	—	Destination conversation
SMTP_HOST, SMTP_PORT, SMTP_USER, SMTP_PASSWORD, SMTP_FROM, SMTP_TO	port 587	Email channel of send_summary

---

Tests

uv sync --group dev
uv run pytest

The suite covers the pure functions (.env resolution, key normalisation, notification dispatch), the geometry of captures on the real screen, and the MCP server contract — its name, its tool list, and the tightness of read-only mode.

Two autouse guardrails are set in tests/conftest.py and checked by tests/test_guardrails.py: httpx and smtplib raise as soon as they are touched. The suite runs on the machine it drives; it must not be able to send a real message.

What is not covered, and why. Input injection — type_text, mouse_click, hotkey — is not tested automatically: a test checking that Ctrl+W works closes a window along the way. Those functions are validated by hand. Likewise, the capture tests assert nothing about the content of the image, only about its structure: nothing may depend on what is displayed at the time of the test.

---

Layout

src/winctl/
  server.py    definition of the 23 MCP tools
  winapi.py    ctypes layer: SendInput, windows, clipboard
  capture.py   screen captures, resizing, coordinate grid
  apps.py      application resolution and launching
  notify.py    Telegram, email, Windows notification, .env loading

capture and notify are not third-party libraries: they are the internals of the server. capture backs the capture tools, notify backs send_summary.

---

License

MIT — see LICENSE.