Summary

A button press is an exact element target by role and name, not a coordinate guess.

That is what survives a layout shift, and it is why a click does not need a screenshot or an LLM call in the loop. Browser work goes through the DevTools Protocol in the background, so Chrome does not need focus — which also means you can keep using your machine while it works.

What it is

A desktop automation MCP server for macOS and Windows. It drives native applications through the accessibility API, and Chrome through the DevTools Protocol in the background.

What you get

111 tools across desktop control, browser automation and fallback execution
Native UI: ui_tree, ui_find, ui_press, ui_set_value, menu_click, click, type_text, key, drag, scroll
Seeing the screen when you need to: screenshot, screenshot_file, ocr, click_text
App handling: apps, windows, focus, launch, and applescript on macOS
Browser over CDP without stealing focus: browser_tabs, browser_open, browser_navigate, browser_js, browser_dom, browser_click, browser_type, browser_wait, browser_page_info
Human-like input and anti-detection: browser_fill_form, browser_human_click, browser_stealth
Automatic fallback from accessibility API to CDP to OCR to coordinates: execution_plan, click_with_fallback, type_with_fallback
Prebuilt app references, playbooks and UI maps that load when the matching app or site is detected

Requirements

macOS or Windows. Add it with claude mcp add screenhand -- npx -y screenhand, or run npx -y screenhand from any MCP client. On macOS, enable your terminal under System Settings > Privacy & Security > Accessibility. Browser tools need Chrome started with --remote-debugging-port=9222. Version 0.5.3.

Setup effort

One command — npx -y screenhand

ScreenHand