mcp-winaccess-win

Version 1.8.0 · Windows-only desktop automation MCP server. Built directly on the OS APIs — ctypes (SendInput, ImageGrab, Win32 windows, clipboard) and UI Automation via comtypes — with no pyautogui / pywinauto / pywin32.

Requirements

• Windows 10/11 (x64) — this package only runs on Windows.

• Python 3.10+.

• The server must run inside an interactive desktop session (a logged-in user), not as a background service — otherwise SendInput/UI Automation cannot reach the desktop.

• No administrator rights are required for normal user applications (see Limitations).

• OCR auto-install needs network access and 7-Zip (or an existing Tesseract install; see Troubleshooting).

Related MCP server: Windows MCP Server

Install

pip install mcp-winaccess-win              # OCR support (pytesseract) is included
pip install "mcp-winaccess-win[vision]"    # optional: OpenCV template matching

The Tesseract binary (a system program, not a Python package) is downloaded and installed automatically on first OCR use into %LOCALAPPDATA%\mcp-winaccess\tesseract (with eng + rus language data). Use --tesseract_cmd <path> to point at an existing install, or --no-auto-tesseract to disable the automatic install (see Troubleshooting).

Run

python -m mcp_winaccess_win.server
# or the console script
mcp-winaccess-win

Command-line arguments

Local (stdio) is the default and needs no arguments. To serve over HTTP, protect the desktop with a token — a token is required whenever --listen is not loopback:

mcp-winaccess-win --transport remote --listen 127.0.0.1 --port 8765 --token mysecret

OpenCode config (opencode.jsonc)

Local (stdio) — OpenCode launches the server itself:

{
  "mcp": {
    "winaccess": {
      "type": "local",
      "command": ["uvx", "mcp-winaccess-win", "--tesseract_cmd", "auto"],
      "enabled": true
    }
  }
}

Remote (streamable-http) — the token goes in the Authorization header:

{
  "mcp": {
    "winaccess": {
      "type": "remote",
      "url": "http://127.0.0.1:8765/mcp",
      "headers": { "Authorization": "Bearer mysecret" },
      "enabled": true
    }
  }
}

Architecture

Modules: mcp_winaccess_win/server.py (tool wrappers), mcp_winaccess_win/adapter/windows.py (adapter), mcp_winaccess_win/adapter/win32_input.py, mcp_winaccess_win/adapter/win32_screen.py, mcp_winaccess_win/adapter/win32_window.py, mcp_winaccess_win/adapter/win32_clipboard.py, mcp_winaccess_win/adapter/win32_overlay.py (element highlight), mcp_winaccess_win/adapter/win32_notifications.py, mcp_winaccess_win/adapter/win32_process.py (launch/kill/enumerate processes), mcp_winaccess_win/adapter/tesseract_setup.py (auto-install OCR engine), mcp_winaccess_win/adapter/uia.py (UI Automation), mcp_winaccess_win/adapter/base.py (interface + shared helpers).

Conventions

• value (window identifier): a partial title/class (str), a window handle (int), or a prefix title:, class:, pid:N, exe:name.exe.

• control_identifier: the ID returned by get_all_controls (element_N) or a control Name/AutoID.

• All coordinates are absolute screen pixels (virtual-desktop space).

• Any failure returns a string starting with ERROR:; success messages are human-readable.

• Tool schemas carry per-parameter descriptions, Literal enums for fixed choices, a human-readable title, and behavioural hints (readOnlyHint/destructiveHint/idempotentHint) so agents can tell observing tools from state-changing ones.

Tools

94 tools. They are registered only when the adapter advertises the matching capability; unsupported tools are hidden.

Screenshots and display — screenshot, screenshot_region, screenshot_monitor, screenshot_window, screenshot_element

• screenshot(grid=False, include_cursor=False) — the whole virtual desktop (all monitors); grid=True overlays labeled 100 px lines; include_cursor=True draws the mouse pointer.

• screenshot_jpg(path="") — save a JPEG (temp file when path is empty).

• screenshot_region(left, top, width, height, grid=False).

• screenshot_monitor(index=0, grid=False) — a single monitor.

• screenshot_window(value, grid=False) — capture a window (even if partially occluded).

• screenshot_element(value, control_identifier, grid=False) — capture a single control.

• compare_screenshots(image_a, image_b, save_diff="") — differing pixels/percentage between two image files.

• list_monitors() (includes DPI scale), image_to_screen_coords(x, y, monitor_index=0), get_pixel_color(x, y), wait_for_pixel_color(x, y, color, timeout=10, tolerance=0).

Mouse and keyboard — input

• get_mouse_position(), move_mouse(x, y), mouse_move_relative(dx, dy).

• click(x=None, y=None, button="left", clicks=1) — omit x/y to click at the current position; clicks=2 is a double-click, button="right" a right-click.

• mouse_down(x=0, y=0, button="left") / mouse_up(button="left") — hold/release for manual drags.

• drag(x_from, y_from, x_to, y_to, duration=0.5), scroll(direction="down", amount=3, x=0, y=0).

• type_text(text) (Unicode), hotkey(keys) (e.g. "ctrl+shift+s").

• press_key(key) — a key (enter, tab, esc, f5, letters/digits) or a system key (win, volumeup, volumedown, volumemute, playpause, nexttrack, prevtrack, printscreen).

• key_down(key) / key_up(key) — hold/release keys (e.g. Shift range selection).

• sleep(seconds) — wait a fixed amount of time.

Clipboard — clipboard

• clipboard_get(), clipboard_set(text), clipboard_clear().

• clipboard_set_files(paths) / clipboard_get_files() — file paths (CF_HDROP).

• clipboard_set_image(path) / clipboard_get_image(save_path="") — image (CF_DIB).

Vision / OCR — vision, ocr

• locate_image_on_screen(image_path, confidence=0.9), wait_for_image(...), click_image(...) — require the vision extra.

• ocr_screen(left, top, width, height, lang=""), find_text_on_screen(text, confidence, lang), click_text(text, confidence, lang) — lang e.g. "eng" or "rus+eng"; the Tesseract binary auto-installs on first use.

• wait_for_text(text, timeout, lang, confidence) — waits until OCR finds text (for canvas/custom UI).

Windows — window_list, window_activate, window_close, window_minimize, window_maximize, window_move, window_resize, window_snap, window_topmost

• list_windows(process=""), find_window(value), get_active_window(), get_window_state(value) (rect, handle, pid, exe, state, monitor).

• switch_to_window(value), wait_for_window(value, timeout, require_ready=False) (require_ready=True waits until visible and responsive), wait_for_window_gone(value, timeout).

• minimize_window(value), maximize_window(value), restore_window(value), close_window(value).

• move_window(value, x, y), resize_window(value, x, y, width, height), snap_window(value, position).

• set_always_on_top(value, enabled).

value accepts a partial window title (str), a window handle (int), or a prefix: title:, class:, pid:N, exe:name.exe. list_windows(process=...) filters by exe name or PID.

Menus and dialogs — menus, dialogs, tray

• menu_select(value, path) (e.g. "File->Save As"), get_menu_items(value, menu="") (discover the menu bar / a menu's items), context_menu_click(x, y, item).

• list_dialogs(), handle_dialog(button_text, title="").

• file_dialog_set_path(path, title=""), file_dialog_confirm(title="").

• tray_icon_click(name, action="left").

UI tree (semantic controls) — ui_tree

• get_all_controls(value, limit=150, query="", control_type="") — use the returned ID (element_N) or Name/AutoID as control_identifier; each row includes Rect and Center.

• click_element(value, control_identifier, clicks=1) (clicks=2 = double-click), get_text, set_text(value, control_identifier, text, mode="value", clear=False) (mode="type" forces focus + typing; clear=True selects-all before typing), get_window_text(value) (flat text dump), select_item, toggle_checkbox, get_control_state, set_slider, get_selected_text, scroll_into_view, wait_for_element, drag_element.

Element inspection — ui_tree

• get_element_at_point(x, y) — element under a screen point.

• get_active_element() — the currently focused control.

• highlight_element(value, control_identifier, duration=1.0, color="red", width=3) — temporary rectangle (visual confirmation).

• highlight_region(x, y, width, height, duration=1.0, color="red", border_width=3) — temporary rectangle over an arbitrary screen region.

• wait_for_element_gone(value, control_identifier, timeout=10) — wait until a control disappears.

• expand_element(value, control_identifier) / collapse_element(value, control_identifier) — tree/expander controls.

• get_control_state(value, control_identifier) — enabled/offscreen/checked/value (works for checkboxes, menu items, sliders).

• drag_element_to_element(from_value, from_control, to_value, to_control, duration=0.5) — drag-and-drop between controls.

• scroll_element(value, control_identifier, direction="down", amount=3).

Lists and tables — ui_tree

• get_list_items(value, control_identifier, limit=200) — item names of a list/tree.

• get_table_data(value, control_identifier, limit=200) — table/grid rows (cells joined by |).

Notifications — notifications

• list_notifications() — current Windows notifications via the UserNotificationListener API (no UI opened).

• dismiss_notifications() — clears all notifications (UI clear-all button; the WinRT clear API is unavailable to desktop apps).

Virtual desktops — desktop

• switch_desktop(direction="right") — switch to the adjacent desktop.

• move_window_to_desktop(value, direction="right") — move a window to the adjacent desktop.

Processes and shell — process, shell

• run_app(path, args="", cwd="") — launch an application detached (returns its PID); pair with wait_for_window(require_ready=True).

• kill_process(pid_or_name) — force-kill a process tree (irreversible; prefer close_window).

• list_processes(filter="") — list running processes as PID N | name.exe.

• run_command(command, cwd="", timeout=30, shell="cmd") — run a console command and return its exit code and output (shell="powershell" for PowerShell). Disable with --no-shell.

Examples

# 12 + 34 in Calculator
click_element("Калькулятор", "clearButton")
click_element("Калькулятор", "num1Button")
click_element("Калькулятор", "num2Button")
click_element("Калькулятор", "plusButton")
click_element("Калькулятор", "num3Button")
click_element("Калькулятор", "num4Button")
click_element("Калькулятор", "equalButton")
get_text("Калькулятор", "CalculatorResults")   # -> 46

# Screenshot with a coordinate grid
screenshot(grid=True)

Limitations

• UIPI: SendInput cannot drive windows with higher integrity (elevated apps) from a non-elevated process. Run the server elevated to automate those.

• Foreground activation: switch_to_window uses the Alt-unlock + SetForegroundWindow technique; some always-on-top/system windows may still refuse focus.

• UWP apps (e.g. Calculator) can expose two ApplicationFrameWindow instances (one minimized); the adapter prefers the visible/largest frame.

• DPI: the process is per-monitor-v2 DPI aware, so coordinates match physical pixels.

Troubleshooting

• OCR fails — the Tesseract binary is auto-installed on first use (downloaded via the UB-Mannheim installer, extracted with 7-Zip, into %LOCALAPPDATA%\mcp-winaccess\tesseract). If that fails (no network / no 7-Zip), install Tesseract manually and pass --tesseract_cmd <path>. Language data eng and rus are fetched automatically; add more by dropping *.traineddata into the managed tessdata.

• Vision tools missing — install the vision extra (opencv-python).

• UI tree tools missing — comtypes is required (installed with the package).

Tests

The tests/ suite was removed; verification is done manually, function by function, against a live desktop session.