cua-driver-install-macos

---

name: cua-driver-install-macos description: Install, configure, and integrate Cua Driver (trycua/cua) on macOS for background desktop control via MCP, including permission granting and Hermes integration. triggers: - install cua - cua setup - cua driver install

---

Cua Driver macOS Install & Integration

When to use

• User wants Hermes to control macOS desktop apps (browser clicks, typing, app launching)
• Need to bypass macOS 26.x restrictions on AppleScript/cliclick/CGEventPost desktop control
• Installing trycua/cua driver for background computer-use

Installation

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

This installs:

• /Applications/CuaDriver.app — the signed macOS bundle
• ~/.local/bin/cua-driver — CLI symlink

Grant Permissions (Critical step)

Run the grant command — this launches CuaDriver so the TCC permission dialogs attach to the correct identity: cua-driver permissions grant

User must manually approve two macOS dialogs:

1. Accessibility (輔助使用) — tick CuaDriver
2. Screen Recording (螢幕錄影) — tick CuaDriver

After toggling Screen Recording, macOS says "Quit and reopen to enable". Restart the daemon: pkill -f CuaDriver then open -n -g -a CuaDriver --args serve

Verify: cua-driver permissions status should show both ✅ granted with source: driver-daemon

Start the Daemon

cua-driver serve & or open -n -g -a CuaDriver --args serve

Verify with cua-driver status.

Set capture mode: cua-driver config set capture_mode som

Generate MCP Config

cua-driver mcp-config --client hermes prints the YAML snippet to add to the agent's config.

Browser Automation via page Tool (Preferred for Web Apps)

For web apps (Gemini, ChatGPT, web forms), AX-based type_text/click often fails because Chromium web content doesn't fully expose AX actions. Use the page tool instead — it works via AppleScript + JavaScript.

Step 0: Enable JavaScript Apple Events (Once)

Chrome blocks JavaScript from Apple Events by default. Enable it:

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

This restarts Chrome. After restart, find the new pid + window_id:

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

Step 1: Navigate

echo '{"pid":PID,"window_id":WID,"action":"execute_javascript","javascript":"window.location.href=\"https://gemini.google.com/app\""}' | cua-driver call page

Step 2: Type into input field via JS

echo '{"pid":PID,"window_id":WID,"action":"execute_javascript","javascript":"(function(){ let el = document.querySelector(\"[contenteditable], textarea, [role=textbox]\"); if(!el) return JSON.stringify({found:false}); if(el.tagName===\"TEXTAREA\"||el.tagName===\"INPUT\"){ el.value=\"your text here\"; el.dispatchEvent(new Event(\"input\",{bubbles:true})); el.dispatchEvent(new Event(\"change\",{bubbles:true})); return JSON.stringify({found:true, tag:el.tagName, value:el.value}); } else { el.innerText=\"your text here\"; el.dispatchEvent(new Event(\"input\",{bubbles:true})); return JSON.stringify({found:true, tag:el.tagName, innerText:el.innerText}); } })()"}' | cua-driver call page

Step 3: Submit (send Enter keystroke via JS)

echo '{"pid":PID,"window_id":WID,"action":"execute_javascript","javascript":"(function(){ let el = document.querySelector(\"[contenteditable], textarea, [role=textbox]\"); if(!el) return \"no input\"; let ev = new KeyboardEvent(\"keydown\", {key:\"Enter\", code:\"Enter\", keyCode:13, which:13, bubbles:true, cancelable:true}); el.dispatchEvent(ev); return \"dispatched Enter\"; })()"}' | cua-driver call page

Step 4: Read response

echo '{"pid":PID,"window_id":WID,"action":"get_text"}' | cua-driver call page

Other page Actions

• query_dom — CSS selector queries
• click_element — click by CSS selector (with cursor animation)
• get_text — extract visible page text

Simple AX-based Workflow (for Native macOS Apps)

1. cua-driver call launch_app '{"bundle_id":"com.google.Chrome","urls":["https://gemini.google.com/app"]}'
2. cua-driver call list_windows '{"pid":PID}' — get window_id
3. cua-driver call get_window_state '{"pid":PID,"window_id":WID}' — snapshot + element indices
4. cua-driver call click '{"pid":PID,"window_id":WID,"element_index":N}' — click
5. cua-driver call type_text '{"pid":PID,"text":"..."}' — type
6. cua-driver call press_key '{"pid":PID,"key":"return"}' — submit

Pitfalls

• "你要還原網頁嗎?" dialog: After opening Chrome, a restore dialog may block the main page. Close it first by clicking the "關閉" button (element_index: 2).
• Element indices are per-snapshot: Always call get_window_state fresh before using an element_index.
• Permissions persist across restarts: As long as the same CuaDriver.app identity is used.
• Screen Recording toggle needs restart: After allowing in System Settings, kill + restart the daemon.
• Background launch: Apps open with self_activation_suppressed=true, meaning they don't pop to front.
• CGEvent fallback: Chromium apps often don't support AXSelectedText, so type_text falls back to CGEvent character synthesis.
• Web apps need page tool, not AX: Gemini/Google/React apps don't respond to AX type_text. Use page tool with JavaScript DOM manipulation instead.
• Quoted JSON on macOS shell: Complex JSON with nested quotes breaks on macOS shell. Always pipe JSON via stdin: echo '{"key":"value"}' | cua-driver call page — avoids PowerShell quote stripping and shell escaping issues.
• page tool needs JS Apple Events enabled: Without this, Chrome returns osascript error 12 ("JavaScript from Apple Events is disabled"). Enable once with enable_javascript_apple_events action — Chrome restarts afterward.
• After Chrome restart from JS Apple Events: Chrome relaunches with a new pid. Find it with list_windows — the Gemini page may open as a new tab, navigate back with window.location.href.
• Gemini Enter submission via JS: Gemini's send button is not a standard <button> — dispatching a KeyboardEvent('keydown', {key:'Enter'}) on the contenteditable div triggers the send. The response appears after ~15-30 seconds of "thinking".