Cua Driver Macos Automation

Cua Driver macOS Desktop Automation

When to use

• User wants to automate macOS desktop apps (browsers, dev tools, etc.)
• Need to click buttons, type text, or interact with web pages programmatically
• macOS native Automation (AppleScript/cliclick) is blocked by OS restrictions
• Want background (non-focus-stealing) desktop control

Version Updates

v0.6.8 (2026-06-28)

• Hermes 原生支援！ 執行 cua-driver mcp-config --client hermes 可直接輸出 Hermes config.yaml
• 更新後要 restart daemon：cua-driver stop && cua-driver serve
• TCC grants 理論上保留，但 macOS 可能出一次性 re-grant prompt
• 如果 permission gate 卡住：cua-driver serve --no-permissions-gate
• 新增 mcp-config 指令，支援 claude/codex/openclaw/cursor/opencode/hermes 等 client
• Update 方法：cua-driver update --apply

Installation

/bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/trycua/cua/main/libs/cua-driver/scripts/install.sh)"

Permissions Setup

After installation, grant two macOS permissions:

1. Accessibility — System Settings > Privacy & Security > Accessibility > tick CuaDriver
2. Screen Recording — System Settings > Privacy & Security > Screen Recording > tick CuaDriver

Check status:

cua-driver permissions status

Trigger permission dialogs:

cua-driver permissions grant

Core Workflow

1. List running apps

cua-driver call list_apps

2. Launch app in background (no focus steal)

cua-driver call launch_app '{"bundle_id":"com.google.Chrome","urls":["https://gemini.google.com/app"]}'

Useful flags:

• creates_new_application_instance: true — force new instance for concurrent sessions
• urls — array of URLs/file paths to open

3. List windows to find window_id

cua-driver call list_windows '{"pid":}'
cua-driver call list_windows '{"on_screen_only":true}'

4. Snapshot + get UI tree (element_index)

cua-driver call get_window_state '{"pid":,"window_id":}'

Use query to filter large trees:

cua-driver call get_window_state '{"pid":,"window_id":,"query":"button|input|search"}'

5. Click by element_index

cua-driver call click '{"pid":,"window_id":,"element_index":}'

6. Type text

cua-driver call type_text '{"pid":,"text":"Hello world"}'

7. Keyboard shortcuts

cua-driver call press_key '{"pid":,"key":"return"}'
cua-driver call hotkey '{"pid":,"keys":["cmd","c"]}'

Web Page Automation (Critical Path)

For interacting with web pages, elementindex + typetext often fail due to web component shadow DOMs.

Enable JavaScript Apple Events (one-time per browser)

echo '{"action":"enable_javascript_apple_events","bundle_id":"com.google.Chrome","user_has_confirmed_enabling":true}' | cua-driver call page

This restarts the browser.

Execute JavaScript directly in the page

Method A — Short JS (inline):

echo '{"pid":,"window_id":,"action":"execute_javascript","javascript":"document.title"}' | cua-driver call page

Method B — Long JS (file → stdin pipe, preferred):

# Write JS to file first, then pipe via python for proper JSON escaping
cat /tmp/script.js | python3 -c "
import sys, json
js = sys.stdin.read()
payload = json.dumps({'pid': , 'window_id': , 'action': 'execute_javascript', 'javascript': js})
with open('/tmp/payload.json', 'w') as f:
    f.write(payload)
" && cat /tmp/payload.json | cua-driver call page

Method C — AppleScript fallback (for Safari when page tool times out):

osascript -e 'tell application "Safari" to do JavaScript (read posix file "/tmp/script.js") in current tab of window 1'

Other page actions

echo '{"pid":,"window_id":,"action":"get_text"}' | cua-driver call page
echo '{"pid":,"window_id":,"action":"query_dom","css_selector":"button"}' | cua-driver call page

React SPA Text Input Strategy (e.g. X.com, Medium)

For React controlled components, standard textContent assignment doesn't trigger React's state update. Use this 3-layer approach:

Layer 1 — execCommand('insertText'): Best for React Draft.js editors (X.com compose box).

const el = document.querySelector('[data-testid="tweetTextarea_0"]');
el.focus();
const sel = window.getSelection();
const range = document.createRange();
range.selectNodeContents(el);
sel.removeAllRanges();
sel.addRange(range);
document.execCommand('insertText', false, tweetText);

Layer 2 — Cua typetext (method='keyevents'): Cua's type_text tool synthesises CGEvent keystrokes. Works when JS execCommand fails but may cause character duplication on SPA inputs. Use delay_ms: 10 for stability.

# Step 1: Click the input area first (use screenshot+vision to find pixel coords)
cua-driver call click '{"pid":,"window_id":,"x":,"y":}'
# Step 2: Type text via keystroke synthesis
cua-driver call type_text '{"pid":,"text":"Your tweet content here","delay_ms":10}'

Layer 3 — ClipboardEvent paste: Dispatches a programmatic paste event that React Draft.js recognizes.

const dataTransfer = new DataTransfer();
dataTransfer.setData('text/plain', tweetText);
el.dispatchEvent(new ClipboardEvent('paste', {
  bubbles: true, cancelable: true, clipboardData: dataTransfer
}));

X.com specific:

• Post button [data-testid="tweetButtonInline"] is disabled until React state has content
• Pixel coordinates for clicking Post button typically ~(610, 420) on 1600×796 Safari window
• After typing via typetext, use screenshot + visionanalyze to verify Post button state
• If Safari shows "確定要離開此網頁嗎?" dialog, click "停留在網頁" to dismiss

Common Pitfalls

1. Shell quoting: When passing JSON with long JS strings, prefer stdin piping.
2. elementindex is per-snapshot: Always call get_window_state before using elementindex.
3. Web apps ignore AX type_text: Chromium/Electron web inputs often don't implement kAXSelectedText.
4. After "Allow JS Apple Events": Chrome restarts with a NEW pid.
5. list_windows can return no windows: Some Electron apps report 0 windows on main PID but have windows on child Renderer process PID.
6. hotkey vs press_key: Use hotkey for keyboard shortcuts (more reliable on Electron apps).
7. Screenshot capture: v0.5.3 get_window_state does NOT include screenshot despite capture_mode. Use macOS screencapture -x /tmp/screen.png as fallback.

Verification

cua-driver status
cua-driver permissions status
cua-driver list-tools

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: ChrisLamDev
• Source: ChrisLamDev/cua-desktop-automation-skills
• License: MIT
• Homepage: https://chrislamdev.github.io/cua-desktop-automation-skills/

Install and usage instructions live in the source repository linked above.