midscene-computer-automation
by quanru
Web & Frontend DevelopmentBrief
Vision-driven desktop automation using Midscene.
Description
What This Skill Does
Controls macOS, Windows, and Linux desktops with natural language using Midscene.js. Operates from screenshots, so it works with any visible UI without requiring DOM access or accessibility labels. Automates clicks, typing, window management, and app interactions across any application.
Unlike Selenium or AppleScript, it works with any app on any stack because it reads the screen directly rather than relying on DOM structure or accessibility APIs.
When to Use It
- Automating repetitive form entry in a legacy desktop app
- Taking targeted screenshots of specific windows for documentation
- Launching apps and navigating their menus via text commands
- Extracting visible on-screen data from any GUI application
- Running multi-step workflows across different desktop programs
Example Workflow
User asks: Open Chrome, search for the weather in New York, and tell me the result
- 1
Connect to the desktop with
npx @midscene/computer@1 connect - 2
Open Chrome using the system launcher (
open -a 'Google Chrome'on macOS) - 3
Take a screenshot to confirm Chrome is in the foreground
- 4
Run
act --prompt "search for weather in New York and read the current conditions from the page"to perform the search and extract results - 5
Disconnect and report the weather information extracted from the screen
Current weather conditions for New York, with a screenshot showing the search result page
Requirements
- MIDSCENE_MODEL_API_KEY from a supported visual AI model provider (Google Gemini, Alibaba Dashscope, Doubao, or OpenRouter)
- MIDSCENE_MODEL_NAME, MIDSCENE_MODEL_BASE_URL, and MIDSCENE_MODEL_FAMILY set to match the chosen provider
Installation
clawhub install quanru/midscene-computer-automationRun the CLI command, or paste the prompt into Claude Code, Cursor, Codex or whichever coding agent you use.