ispyx

A local desktop-control MCP server. It lets an AI assistant use your computer the way a person does: look at the screen, move the mouse, click, drag, scroll, type, press hotkeys and use the clipboard.

• Runs on your machine, talks to the AI app over stdio, opens no network port (the settings panel only listens while it is open).

• Windows and macOS. Windows needs Node.js 22+; the macOS app carries its own.

• One resident app (tray icon on Windows, menu-bar icon on macOS) does the work. Every AI app connects through a thin ispyx mcp process, so there is one stop key, one banner and one set of settings.

Install

Windows

Open PowerShell and paste:

irm https://raw.githubusercontent.com/murat-mehmet/ispyx/master/scripts/install.ps1 | iex

No administrator rights needed. It finds or installs Node.js 22+ for your user, runs npm install -g ispyx, adds an Ispyx shortcut to your Desktop, turns on start at login and starts ispyx. Manual: install Node.js from nodejs.org, then npm i -g ispyx and run ispyx.

Uninstall: run the installer with -Uninstall (add -Purge to also delete settings and logs). Options and troubleshooting: docs/windows-install.md.

macOS

Open Terminal and paste:

curl -fsSL https://github.com/murat-mehmet/ispyx/releases/latest/download/install.sh | sh

It downloads the newest Ispyx.app, checks it (SHA-256, the developer's signature, notarised) and puts it in /Applications, then opens it; allow the two permissions its window asks for. Or download the .dmg from the Releases page and drag Ispyx to Applications. The app carries its own Node, nothing else is needed. Uninstall: ... | sh -s -- --uninstall (add --purge for settings and logs). Details: native/macos/README.md.

Updating

The panel shows when a newer version is out, with an Update button. Or run ispyx update (macOS: /Applications/Ispyx.app/Contents/Resources/ispyx update, or the install command above again). Windows reinstalls the npm package; macOS replaces the app, which keeps its permissions. Settings and logs stay.

Related MCP server: vision-input

Connect your AI app

Open the panel (tray or menu-bar icon, Open): it shows the exact command for your machine with a copy button. Or by hand:

claude mcp add ispyx --scope user -- ispyx mcp

Other apps (Claude Desktop, LM Studio, ...), JSON form. On Windows some apps need "command": "cmd", "args": ["/c", "ispyx", "mcp"]:

{
  "mcpServers": {
    "ispyx": { "command": "ispyx", "args": ["mcp"] }
  }
}

On macOS use /Applications/Ispyx.app/Contents/Resources/ispyx as the command, without mcp. The AI app starts ispyx when needed, so start at login is optional.

Using it

• Panel. Permissions (macOS), recent activity, effects, password, denylist, start at login, the connect command. Opens in its own small window and closes when you close it.

• Commands. ispyx starts the resident app and opens the panel, ispyx mcp is the stdio server for an AI app, ispyx quit stops it, ispyx update installs the latest version and restarts it. On macOS they are /Applications/Ispyx.app/Contents/Resources/ispyx update / quit (the app starts the resident itself). Errors go to %APPDATA%\ispyx\ispyx.log (macOS: ~/Library/Application Support/Ispyx/ispyx.log). Details: src/core/RESIDENT.md.

• Stop. Alt+Esc (Option+Esc on macOS): one press cancels everything running, a second press within 5 s shuts the server down.

• Feedback. A banner shows while a tool runs (green working, amber cancelled, red stopped). An eye and a scan line show on screenshots, clicks, scrolls and typing get small animations. All of it is click-through and can be turned off in the panel.

Tools

Typical flow: screenshot_screen, read coordinates off the image, click (with space: "shot") or type_text, then screenshot_screen to check. Coordinates and per-tool behaviour: docs/behavior.md. Input is human-paced: the pointer glides along a curved path, scrolling is smoothed, typing has natural pauses.

Safety

• Alt+Esc is caught by a low-level keyboard hook, so it does not also cycle windows.

• Denylist of programs that are never focused or driven (password managers by default); editable in the panel.

• No credentials. Typing into password fields is refused; ispyx never reads or types saved passwords.

• Password gate (optional, set it in the panel): asks before the first tool call.

• Action log. Every action, refusal and error is logged locally, including typed text (not text aimed at a password field). Clipboard reads log the length only.

• Elevated windows cannot be controlled on Windows (UIPI), and neither can UAC prompts or the lock screen.

• This gives an AI control of your computer. Connect it only to apps you trust.

Configuration

Most settings are in the panel. Files and variables, later wins: built-in defaults, config.json (package folder or the file in ISPYX_CONFIG), the per-user %APPDATA%\ispyx\config.json (macOS: ~/Library/Application Support/Ispyx/), environment variables. See config.example.json.

More keys (maxTypeLength, moveMinMs/moveMaxMs, typeMinMs/typeMaxMs/typeCapMs) are in config.example.json. The older WIN_INPUT_* names still work as fallbacks.

Development

git clone https://github.com/murat-mehmet/ispyx
cd ispyx
npm install
npm test        # unit checks; other scripts in docs/testing.md

• src/core/ shared: MCP server and tools, safety guard, input motion, resident app, panel, password gate.

• src/platform/ the OS layer (contract in src/platform/CONTRACT.md): win32/ (Win32 through koffi), darwin/ (a client of Ispyx.app).

• src/effects/ animations. native/macos/ the macOS app (Swift). scripts/install.ps1 the Windows installer.

• Releases (npm package and macOS DMG) are made by pushing a version tag: docs/releasing.md.

Tests drive the real desktop and use their own settings (own stop chord, own pipe), so they never touch a server you are using.

License

See LICENSE.