#Desktop

Midscene uses native keyboard and mouse controls to automate desktop applications on Windows, macOS, and Linux.

This guide covers platform setup, model configuration, Playground, and JavaScript SDK integration with @midscene/computer.

#See it in action

Prompt (macOS): Open Safari and post a tweet announcing that Midscene supports AutoGLM. Use the AutoGLM video from the Downloads folder.

View the full report, or explore more Midscene showcases.

#Key features

Cross-platform desktop control

• Mouse operations: click, double-click, right-click, move, and drag-and-drop.
• Keyboard input: type text and press keys with Cmd, Ctrl, Alt, or Shift.
• Screen capture: take screenshots from any display.
• Multiple displays: work with multiple monitors in one workflow.

AI-powered automation

Use natural-language instructions to automate desktop applications:

await agent.aiAct('open the File menu');
await agent.aiAct('click on Save As');
await agent.aiAct('type "my-document" in the filename field');
await agent.aiAct('press Enter');

#Use cases

• Test Electron, Qt, and native desktop applications.
• Automate repetitive work across desktop applications.
• Control multiple applications in sequence.
• Test desktop interfaces with natural-language instructions.

#Get started

#Prepare your desktop environment

#Node.js

Node.js 18.19.0 or higher is required.

#Platform-specific dependencies

macOS: Accessibility permissions are required for keyboard and mouse control. When you run the script for the first time, macOS will prompt you to grant access. Go to System Settings > Privacy & Security > Accessibility and enable permissions for the application running your script (e.g., Terminal, iTerm2, VS Code, WebStorm, or other IDEs). For more details, see nut.js macOS setup.

Windows: No extra setup is needed for ordinary apps. However, Windows isolates input across privilege levels (UIPI): a non-elevated process cannot send mouse or keyboard input to a window that runs as Administrator (elevated). The input is silently dropped — the cursor still moves to the right spot, but clicks and keystrokes have no effect. Prefer running the target application without Administrator privileges. If the target application must stay elevated, run the terminal or Node.js that launches Midscene as Administrator too, so both processes share the same privilege level. See Windows: clicks have no effect on some apps.

Linux: ImageMagick is required for screenshot functionality.

Headless Linux (CI): To run desktop automation on a headless Linux server (e.g. GitHub Actions), install Xvfb and its dependencies, then enable headless mode:

# Install dependencies
sudo apt-get install -y xvfb x11-xserver-utils imagemagick

// Option 1: Pass headless option
const agent = await agentForComputer({ headless: true });

// Option 2: Set environment variable
// MIDSCENE_COMPUTER_HEADLESS_LINUX=true npx tsx example.ts

Xvfb creates a virtual display so that mouse, keyboard, and screenshot operations work without a physical monitor. See API Reference for details.

#Launch Playground

Playground is the fastest way to validate the connection and try core capabilities such as aiAct, aiQuery, and aiAssert without writing code. It shares the same core as @midscene/computer, so anything that works here will behave the same once scripted.

1. Launch the Playground CLI:

npx --yes @midscene/computer-playground

2. Click the gear icon in the Playground window, then paste your API key configuration. Refer back to Model configuration if you still need credentials.

#Use the JavaScript SDK

Once Playground works, move to a repeatable script with the JavaScript SDK.

#Configure the model

Set the model configuration through environment variables. See Model strategy for guidance on choosing a model.

export MIDSCENE_MODEL_BASE_URL="https://replace-with-your-model-service-url/v1"
export MIDSCENE_MODEL_API_KEY="replace-with-your-api-key"
export MIDSCENE_MODEL_NAME="replace-with-your-model-name"
export MIDSCENE_MODEL_FAMILY="replace-with-your-model-family"

For all configuration options, see Model configuration.

#Install dependencies

npm install @midscene/computer

yarn add @midscene/computer

pnpm add @midscene/computer

bun add @midscene/computer

deno add npm:@midscene/computer

#Write a script

Create example.ts:

import { agentForComputer } from '@midscene/computer';

(async () => {
  // Create an agent
  const agent = await agentForComputer({
    aiActionContext: 'You are controlling a desktop computer.',
  });

  // Take a screenshot and query information
  const screenInfo = await agent.aiQuery(
    '{width: number, height: number}, get screen resolution'
  );
  console.log('Screen resolution:', screenInfo);

  // Move mouse to center
  await agent.aiAct('move mouse to center of screen');

  // Assert screen has content
  await agent.aiAssert('The screen has visible content');

  console.log('Desktop automation completed!');
})();

#Run the script

npx tsx example.ts

#Connect to a Remote Windows Desktop via RDP

@midscene/computer can also drive a remote Windows desktop directly over the RDP protocol through the dedicated agentForRDPComputer() factory.

#Prerequisites

1. A reachable Windows machine with RDP enabled.
2. FreeRDP installed on the machine running your script.

#Example

import { agentForRDPComputer } from '@midscene/computer';

const agent = await agentForRDPComputer({
  aiActionContext:
    'You are controlling a remote Windows desktop over the RDP protocol.',
  host: '10.75.166.249',
  port: 3389,
  username: 'Admin',
  password: 'replace-with-your-password',
  ignoreCertificate: true,
});

await agent.aiWaitFor('The remote Windows desktop is visible');
await agent.aiAct('Click the Windows Start button');
await agent.aiAct('Open Settings');
await agent.aiAssert('The Windows Settings window is visible');

#Common RDP Options

• host: Remote Windows host or IP.
• port: RDP port. Defaults to 3389.
• username / password: Account credentials for the remote session.
• domain: Optional Windows domain.
• ignoreCertificate: Skip certificate validation for self-signed setups.
• desktopWidth / desktopHeight: Request a specific remote desktop resolution.
• adminSession: Request the remote admin session when the server allows it.

RDP sessions are exposed to Midscene as a single remote display. You can still use the same aiAct, aiQuery, aiAssert, and report features as local desktop automation.

#Multi-Display Support

If you have multiple displays, you can control a specific one:

import { ComputerDevice, agentForComputer } from '@midscene/computer';

// List all displays
const displays = await ComputerDevice.listDisplays();
console.log('Available displays:', displays);

// Connect to a specific display
const agent = await agentForComputer({
  displayId: displays[0].id,
});

#Example Usage

#Basic Mouse Operations

// Click at center of screen
await agent.aiAct('click mouse at center of screen');

// Move mouse to a specific location
await agent.aiAct('move mouse to top-left corner');

// Double-click
await agent.aiAct('double-click on the desktop icon');

// Right-click
await agent.aiAct('right-click to open context menu');

#Keyboard Operations

// Type text
await agent.aiAct('type "Hello World"');

// Press keyboard shortcuts
if (process.platform === 'darwin') {
  await agent.aiAct('press Cmd+Space to open Spotlight');
  await agent.aiAct('type "Calculator" and press Enter');
} else {
  await agent.aiAct('press Windows key');
  await agent.aiAct('type "Calculator" and press Enter');
}

// Press function keys
await agent.aiAct('press Escape');
await agent.aiAct('press Enter');

#Query Information

// Extract screen information
const info = await agent.aiQuery(
  '{hasDesktop: boolean, visibleApps: string[]}, check if desktop is visible and list visible apps'
);

// Locate elements
const position = await agent.aiLocate('the File menu');
console.log('File menu position:', position);

#Complex Workflows

// Open an application and interact with it
await agent.aiAct('open Finder');
await agent.aiWaitFor('Finder window is visible');

await agent.aiAct('click on Documents folder');
await agent.aiAct('press Cmd+N to create new folder');
await agent.aiAct('type "My Project"');
await agent.aiAct('press Enter');

await agent.aiAssert('A folder named "My Project" exists');

#Environment Check

You can check if your system is properly configured:

import { checkComputerEnvironment } from '@midscene/computer';

const env = await checkComputerEnvironment();
console.log('Platform:', env.platform);
console.log('Available:', env.available);
console.log('Displays:', env.displays);

if (!env.available) {
  console.error('Environment not available:', env.error);
}

#FAQ

#macOS: Script cannot control mouse or keyboard

macOS requires Accessibility permissions for keyboard and mouse control. Go to System Settings > Privacy & Security > Accessibility and enable the toggle for the application running your script (e.g., Terminal, iTerm2, VS Code, or WebStorm).

If you have already granted permission but it still doesn't work, try removing the app from the Accessibility list and re-adding it — macOS sometimes caches stale permissions.

#Windows: clicks have no effect on some apps

If the cursor moves to the correct position but clicks or key presses do nothing on a particular application — while other apps work fine — check whether the target app is running as Administrator (elevated). Windows UIPI blocks input injected from a lower-privilege process into an elevated window and drops it silently, with no error.

Prefer lowering the target application's privilege level first, for example by launching it without "Run as Administrator" or disabling any setting that always starts it elevated. If the target app must stay elevated, run the terminal or Node.js that launches Midscene as Administrator so it matches the target app's privilege level, then try again. System-level shortcuts such as Win+Tab are handled by the shell and keep working even when this happens, which is why keyboard shortcuts may appear to work while in-app clicks do not.

The health check logged at connection time prints this troubleshooting link when Midscene is not running as Administrator on Windows.

#Linux: Screenshots or interactions fail on a headless server

A headless Linux environment (e.g. CI) has no physical display. You need to install Xvfb and ImageMagick, and enable headless mode:

sudo apt-get install -y xvfb x11-xserver-utils imagemagick

const agent = await agentForComputer({ headless: true });

Or set the environment variable:

MIDSCENE_COMPUTER_HEADLESS_LINUX=true npx tsx example.ts

#Next Steps

• API Reference
• Use YAML format automation scripts
• YAML script runner
• Caching for efficiency
• Demo projects
  • JavaScript SDK demo
  • Vitest demo
  • Remote Windows desktop over RDP
  • Obsidian on headless Linux CI