computer-automation
Vision-driven desktop automation using Midscene. Control your local desktop (macOS, Windows, Linux) or a remote Windows desktop over RDP with natural language commands. Operates entirely from screenshots — no DOM or accessibility labels required. Can interact with all visible elements on screen regardless of technology stack. ⚠️ In local mode this takes over the user's real mouse and keyboard. For web apps, prefer "Browser Automation" instead. Only use this for desktop-native apps (Electron, Qt
Security Assessment
Detected risks:
About computer-automation
Computer Automation is a vision-driven desktop-automation skill built on Midscene that lets Claude control a local desktop (macOS, Windows, Linux) or a remote Windows desktop over RDP using natural-language commands. It operates entirely from screenshots with no DOM or accessibility labels, so it can interact with any visible on-screen element regardless of the underlying technology stack. It solves automation of desktop-native applications (Electron, Qt, native apps) that cannot run in a browser, and explicitly steers web-app tasks to a separate Browser Automation skill instead.
The skill drives the @midscene/computer CLI, where the agent runs one command at a time synchronously, reads the resulting screenshot, and decides the next action in a screenshot-analyze-act loop. An act call can move the mouse, click, double/right-click, drag, type or clear text, scroll, and press keys or shortcuts on the selected display. It requires a vision-capable model configured through environment variables (API key, model name, base URL, family) with examples for Gemini, Qwen, Doubao, and others, and Midscene auto-loads a local .env. RDP mode is activated by passing a host, username, and password to connect, routing all subsequent actions through a bundled RDP helper; supported flags cover port, domain, security protocol, an option to skip TLS certificate validation for trusted dev hosts with self-signed certs, admin-session attachment, and requested desktop resolution. The skill's critical rules require synchronous single-command execution, minimizing rather than closing windows, and always reporting results to the user.
It targets QA engineers, RPA builders, and power users automating desktop-native apps or remote Windows hosts. This is high-capability tooling: in local mode it takes over the user's real mouse and keyboard, and it can drive remote machines over RDP with supplied credentials — genuine capability that warrants care. Mitigating factors are that credential handling follows good practice (read passwords from environment variables or a secrets manager, never paste them into a shared transcript), the certificate-skip option is explicitly limited to trusted dev hosts, and the workflow rules avoid closing the user's applications; there is no exfiltration, persistence, or malicious behavior in the content.
FAQ
What can it control?
A local macOS, Windows, or Linux desktop, or a remote Windows desktop over RDP, interacting with any visible on-screen element from screenshots alone via mouse, keyboard, drag, scroll, and key/shortcut actions.
What does it require to run?
The @midscene/computer CLI (run with npx) and a vision-capable model configured through environment variables — API key, model name, base URL, and family — with examples for Gemini, Qwen, and Doubao; Midscene auto-loads a local .env.
Does it take over my real mouse and keyboard?
In local mode, yes — the skill warns that it takes over the user's real mouse and keyboard, and it advises using Browser Automation for web apps and reserving this for desktop-native apps or remote RDP hosts.
How are RDP credentials handled?
RDP mode is entered by passing host, username, and password to connect; the skill advises reading the password from an environment variable, secrets manager, or interactive prompt and never pasting it into a shared transcript, and it drives the remote host without touching the local mouse or keyboard.
What safety rules does it enforce?
Run one command at a time synchronously so each screenshot can be read before the next action, never run commands in the background, only minimize windows rather than close them unless explicitly asked, and always report results to the user.
Install computer-automation
Quick Setup:
- Copy the skill folder to
.claude/skills/ - Claude will automatically detect and use the skill
Repository
web-infra-dev/midscene-skills