MacOS Desktop Control

Generic macOS desktop control using AppleScript for app and window semantics plus screenshot, OCR, mouse, and keyboard workflows.

写在前面

特别做了中文兼容，包括文字输入/识别等，中文用户放心使用～

macos-desktop-control

This skill controls the macOS desktop through a small, explicit pipeline with a clear split between semantic app control and visual UI control:

Features

🖥️ App and window control

• ✅ Activate an app by name or bundle path
• ✅ Check whether an app is running
• ✅ Read the current frontmost app
• ✅ Read front window title, count windows, and list window titles

📸 Screenshot and image operations

• ✅ Capture the current screen as a logical-resolution screenshot
• ✅ Initialize screenshot-to-click calibration for macOS Retina displays
• ✅ Crop a known rectangular region from an image
• ✅ Reuse calibration data when a workflow must mix logical and raw screenshots

🎯 Visual target location

• ✅ Locate targets by AI semantic understanding as the default first choice
• ✅ Fall back to OCR when the target is best identified by text
• ✅ Fall back to OpenCV image matching when the target has a stable reusable template
• ✅ Constrain later actions to coordinates derived from a screenshot

⌨️ Mouse and keyboard control

• ✅ Move the mouse in logical screen coordinates
• ✅ Left click, right click, double click, and drag
• ✅ Read current mouse position
• ✅ Type text, paste via higher-level workflows, press keys, and send hotkeys
• ✅ Hold and release keys explicitly when needed

🛡️ Safety and scope

• ✅ Use logical coordinates as the default working convention
• ✅ Keep app-specific UI semantics out of this skill
• ✅ Keep AppleScript usage limited to app and window semantics, not deep UI scripting
• ✅ Keep pyautogui.FAILSAFE = True so moving to the top-left corner aborts automation

1. Use AppleScript for app and window semantics
2. Initialize coordinate mapping
3. Capture the screen
4. Locate targets by AI semantic understanding first, then fall back to OCR or OpenCV when needed
5. Execute mouse and keyboard actions with Python

Design boundary

This skill intentionally does not include AppleScript UI scripting.

Use AppleScript for:

• opening or activating apps
• reading frontmost app state
• reading window titles and counts

Use screenshot-guided OCR/OpenCV plus pyautogui for:

• clicking UI targets
• typing into custom-drawn interfaces
• interacting with chat rows, images, canvases, or other visually defined targets

This boundary keeps the skill predictable. AppleScript is used where semantic macOS state is strong, and pyautogui is used where direct UI manipulation is more reliable.

Why initialization is needed

On macOS, screenshot coordinates and click coordinates may use different coordinate systems.

• screencapture images usually use pixel coordinates.
• Mouse automation tools often use macOS screen coordinates, also called point coordinates.
• On Retina displays, one point is commonly equal to two pixels.

This skill writes the coordinate mapping result to a JSON file, so later steps can reuse it without recalculating.

Initialization behavior in the current version:

• the skill auto-initializes on first use when the calibration file does not exist
• it does not re-run mapping on every invocation
• if /tmp/macos_desktop_control/calibration.json already exists, the existing calibration is reused

Default calibration file:

/tmp/macos_desktop_control/calibration.json

Directory layout

macos-desktop-control/
  SKILL.md
  requirements.txt
  scripts/
    calibration.py
    init_coordinate_mapping.py
    capture_screen.py
    crop_image.py
    locate_text_ocr.py
    locate_image_opencv.py
    mouse.py
    keyboard.py
    applescript_app.py
    applescript_window.py

Requirements

Install Python dependencies:

pip install -r requirements.txt

OCR uses Apple Vision through PyObjC, so no separate Tesseract install is required.

On macOS, grant the terminal or runtime app these permissions:

• Screen Recording
• Accessibility

1. Initialize coordinate mapping

The first version handles Retina screens by comparing screenshot pixel size with the logical screen size used by pyautogui.

You can still run initialization manually:

python scripts/init_coordinate_mapping.py

But in normal use, the skill now performs lazy initialization automatically on first use if the calibration file is missing.

Example output:

{
  "screen_width_points": 1512,
  "screen_height_points": 982,
  "screenshot_width_pixels": 3024,
  "screenshot_height_pixels": 1964,
  "scale_x": 2.0,
  "scale_y": 2.0,
  "mode": "retina"
}

Later scripts read this file automatically.

Current lazy-init behavior:

• capture_screen.py
• mouse.py
• locate_text_ocr.py
• locate_image_opencv.py

These scripts first check whether /tmp/macos_desktop_control/calibration.json exists. If not, they auto-generate it once and then continue.

2. Capture screen

Capture the current screen and resize the image into the logical coordinate system used by pyautogui.position() and pyautogui.click().

This skill's default convention is:

• default screenshot is logical
• default recognition result coordinates are logical
• default mouse action coordinates are logical
• default crop operations should use a logical screenshot
• only use calibration conversion when a workflow explicitly mixes logical screenshots with raw pixel screenshots

python scripts/capture_screen.py --output /tmp/macos_desktop_control/screen_logical.png

Core idea:

import pyautogui

img = pyautogui.screenshot()
screen_w, screen_h = pyautogui.size()

# Resize screenshot to the coordinate system used by pyautogui.position() / click().
img = img.resize((screen_w, screen_h))
img.save("screen_logical.png")

3. Crop image regions

When a higher-level skill already knows a target rectangle, crop it directly instead of re-opening previews or re-running visual search.

By default, crop from a logical screenshot so the crop rectangle stays in the same coordinate system as recognition and mouse targeting. Only crop from a raw Retina or pixel screenshot when there is a specific reason to preserve raw pixels, and in that case convert coordinates first using calibration data.

python scripts/crop_image.py \
  --image /tmp/macos_desktop_control/screen_logical.png \
  --x1 400 --y1 300 --x2 700 --y2 650 \
  --output /tmp/macos_desktop_control/crop.png

Use this for:

• extracting a detected chat image thumbnail
• saving a button or dialog region for later analysis
• debugging screenshot-to-action pipelines

4. Locate targets

Use AI semantic understanding as the default first-choice locator. Use OCR or OpenCV only when AI is unsuitable, unavailable, or cannot produce a reliable target.

There are three supported strategies.

Locate by semantic understanding

python scripts/locate_text_ocr.py \
  --image /tmp/macos_desktop_control/screen_logical.png \
  --text "Confirm"

You can also constrain OCR to a specific screen region when the same text may appear in multiple places:

python scripts/locate_text_ocr.py \
  --image /tmp/macos_desktop_control/screen_logical.png \
  --text "Chats" \
  --x1 0 --y1 120 --x2 520 --y2 1107

The script prints the center point of the best matched Apple Vision OCR box. When a region is provided, the search runs only inside that rectangle, but the returned coordinates are still in full-screen logical coordinates.

When exact matching methods are unnecessary or brittle, use AI-based image understanding first.

This approach is the default when the target has one or more of these properties: its appearance is not fixed, there is no reusable template, the decision depends on surrounding visual context, or the target can only be described semantically.

Typical examples:

• there are multiple clickable regions on screen and the caller must determine which one is the intended target
• the target has no stable text label or icon template, but it can be described in natural language
• the workflow requires understanding relationships between visual elements, such as who is speaking or what a region means in context
• the target is easier to describe than to match exactly

Preferred flow:

• capture a screenshot
• crop to the smallest reliable region when possible
• run AI recognition on that region first
• require structured output from AI before acting
• if AI cannot produce a reliable target, fall back to OCR or OpenCV
• map coordinates, then execute the action

AI output contract:

• require a compact JSON object
• required fields: found, x, y, confidence, reason
• found must be boolean
• x and y must be logical screen coordinates when found=true
• confidence should use a small fixed set such as high, medium, low
• reason should briefly explain why the target was selected or why no reliable target was found

Recommended JSON shape:

{
  "found": true,
  "x": 742,
  "y": 681,
  "confidence": "high",
  "reason": "Located the send button in the lower-right input area"
}

Safety rule for AI-driven clicks:

• if found=false, do not click
• if confidence=low, prefer fallback or verification before clicking

This skill is responsible for screenshot capture and coordinate conversion. The caller interprets the recognition result and decides the next action.

Locate by OCR text

5. Mouse actions

Use Python and pyautogui to control the mouse in logical screen coordinates.

Single click

python scripts/mouse.py --action click --x 500 --y 300

Move only

python scripts/mouse.py --action move --x 500 --y 300 --duration 0.2

Double click

python scripts/mouse.py --action double-click --x 500 --y 300

Right click

python scripts/mouse.py --action right-click --x 500 --y 300

Drag

python scripts/mouse.py --action drag --x 500 --y 300 --to-x 800 --to-y 500 --duration 0.3

Read current mouse position

python scripts/mouse.py --action position

You can also pipe the result from a locate script:

python scripts/locate_image_opencv.py \
  --image /tmp/macos_desktop_control/screen_logical.png \
  --template ./target_button.png \
| python scripts/mouse.py --stdin --action click

Stdin accepts either x y text or JSON like {"x": 500, "y": 300}.

6. Keyboard actions

Use Python and pyautogui to paste text or trigger shortcuts.

Important practical note:

• this skill uses clipboard paste for all text entry by default, including English
• this avoids input-method issues with Chinese, English, and mixed-language text
• do not use simulated typing for text entry in this skill

Paste text

python scripts/keyboard.py --action paste --text "I am OpenClaw"

Paste from stdin

printf 'I am OpenClaw' | python scripts/keyboard.py --action paste --stdin

Default input rule for this skill:

• use clipboard paste for all text input by default, including English
• click the verified input field first, then paste with command v
• do not use simulated typing for text entry in this skill

Press one key

python scripts/keyboard.py --action press --key enter

Press a hotkey

python scripts/keyboard.py --action hotkey --keys command v

Recommended paste workflow when text fidelity matters:

1. copy the exact text into the clipboard, preferably via python scripts/keyboard.py --action paste
2. click the verified input field
3. let the script send command v to paste
4. verify visually before pressing enter if sending would be externally visible

Hold and release keys

python scripts/keyboard.py --action key-down --key shift
python scripts/keyboard.py --action key-up --key shift

7. AppleScript app control

Use AppleScript when the task is semantic macOS control rather than visual targeting.

Good fits:

• open or activate an app
• check whether an app is running
• read the current frontmost app

Open by app name

python scripts/applescript_app.py --action open --app "WeChat"

Open by bundle path

python scripts/applescript_app.py --action open --path "/Applications/WeChat.app"

Activate an app

python scripts/applescript_app.py --action activate --app "WeChat"

Check whether an app is running

python scripts/applescript_app.py --action is-running --app "WeChat"

Get the current frontmost app

python scripts/applescript_app.py --action frontmost-app
python scripts/applescript_app.py --action frontmost-app --json-pretty

8. AppleScript window inspection

Use AppleScript window inspection when you need app-level UI state without relying on OCR.

Good fits:

• read the front window title
• count windows for a process
• list window titles for a process

Read the front window title

python scripts/applescript_window.py --action title --app "WeChat"

Count windows

python scripts/applescript_window.py --action count --app "WeChat"

List window titles

python scripts/applescript_window.py --action list --app "WeChat"
python scripts/applescript_window.py --action title --app "WeChat" --json-pretty

Locate by OpenCV image matching

python scripts/locate_image_opencv.py \
  --image /tmp/macos_desktop_control/screen_logical.png \
  --template ./target_button.png \
  --threshold 0.8

The script prints the center point of the matched template.

Use this as a fallback when AI is not appropriate and the target has a stable reusable visual template.

9. When to use AppleScript vs desktop vision

Prefer AppleScript for:

• opening or activating apps
• reading window titles
• checking the frontmost app
• simple app and process state queries

Do not add AppleScript UI scripting here for button clicks or deep accessibility-tree automation. That path is intentionally excluded from this skill.

Prefer screenshot + desktop vision + pyautogui for:

• buttons or labels that only exist visually
• apps with weak or unstable accessibility hierarchies
• targets inside custom-drawn UIs such as chat rows, images, or canvas content
• direct manipulation such as clicking, dragging, and typing into app surfaces

Default visual targeting order in this skill:

1. Try AI semantic understanding first
2. If AI cannot produce a reliable target, fall back to OCR for text-driven targets
3. If OCR is not suitable, fall back to OpenCV template matching for stable visual templates

When the same text may appear in multiple places, do not search the full screen by default. Constrain OCR to the intended region first, then click using the returned full-screen logical coordinates.

A practical sequence is often:

1. AppleScript activates the app
2. AppleScript reads window or process state
3. AI-based vision tries to find the target first
4. OCR or OpenCV is used only as fallback when needed
5. mouse or keyboard automation performs the action
6. AppleScript or a fresh screenshot verifies the result

10. Recommended flow

python scripts/applescript_app.py --action activate --app "WeChat"
python scripts/applescript_window.py --action title --app "WeChat"
python scripts/init_coordinate_mapping.py
python scripts/capture_screen.py
# first try AI semantic understanding with a bounded screenshot region when possible
# if AI cannot produce a reliable target, fall back to OCR or OpenCV
python scripts/locate_text_ocr.py --text "Confirm"
python scripts/mouse.py --action click --x 500 --y 300
python scripts/keyboard.py --action press --key enter

Notes

• Version 1 assumes a Retina display and single primary screen.
• Treat logical screenshots as the default working surface for this skill.
• Treat recognition output coordinates as logical unless a script explicitly says otherwise.
• Treat mouse and keyboard targeting as logical by default.
• Treat crop rectangles as logical by default, and prefer cropping from a logical screenshot.
• If another skill mixes logical screenshots with raw Retina or pixel screenshots, use calibration conversion deliberately. Do not assume logical bounds match raw pixel bounds 1:1.
• Keep this skill focused on generic desktop primitives. App-specific UI semantics, business rules, and event pipelines should stay in the higher-level app skill.
• Treat AI semantic understanding as the default visual locator, not merely a last-resort add-on.
• Require structured AI output with found, x, y, confidence, and reason before acting on AI-located targets.
• Use OCR and OpenCV as fallback tools when AI cannot reliably identify the target.
• All click, drag, move, and typing actions use Python / pyautogui.
• AppleScript support in this skill is limited to app control and window inspection.
• For safety, keep pyautogui.FAILSAFE = True; moving the mouse to the top-left corner aborts automation.