Computer Use Mcp — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited Computer Use Mcp (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
An MCP server that gives AI assistants full macOS desktop control: screenshots, mouse, keyboard, scrolling, and app management. Should work with any MCP client - Claude Code, OpenCode, or your own agent.
Requires Claude.app — the server loads two native binaries bundled inside /Applications/Claude.app. You don't need to be running Claude Desktop to use this server, but the app must be installed.The server exposes 24 tools covering everything needed to operate a macOS desktop:
| Category | Tools |
|---|---|
| Vision | screenshot, zoom |
| Mouse | left_click, right_click, middle_click, double_click, triple_click, mouse_move, left_click_drag, left_mouse_down, left_mouse_up, scroll |
| Keyboard | key, hold_key, type |
| Clipboard | read_clipboard, write_clipboard |
| Apps | request_access, open_application, list_granted_applications, switch_display |
| Utility | cursor_position, wait, computer_batch |
Screenshots are captured at full Retina resolution and scaled to fit model constraints (≤1568px, ≤1.15MP). Click coordinates from the model are automatically mapped back to logical screen coordinates for CGEvent dispatch.
App access is tiered by category, matching the behaviour of Claude Code's built-in computer use:
| Tier | Applies to | What the model can do |
|---|---|---|
| View-only | Browsers (Safari, Chrome, Firefox, Edge, Arc, Brave…), trading platforms | Screenshot only |
| Click-only | Terminals (Terminal, iTerm, Ghostty, Warp), IDEs (VS Code, Cursor, JetBrains…) | Click and scroll, no typing |
| Full control | Everything else | All actions |
/Applications/Claude.appgit clone https://github.com/NicolaivdSmagt/computer-use-mcp.git
cd computer-use-mcp
npm installAdd the server to ~/.claude.json:
{
"mcpServers": {
"computer-use": {
"command": "node",
"args": ["/absolute/path/to/computer-use-mcp/index.js"]
}
}
}Restart Claude Code. The tools appear as mcp__computer-use__* in your session.
Add to your OpenCode config (~/.config/opencode/config.json or equivalent):
{
"mcp": {
"servers": {
"computer-use": {
"command": "node",
"args": ["/absolute/path/to/computer-use-mcp/index.js"]
}
}
}
}The server checks both permissions at startup and exits with a clear error if either is missing.
Permissions are inherited by child processes, so you only need to grant them to the parent process once.
Every session starts with an empty allowlist. Call request_access first with the apps you need:
request_access(apps: ["Safari"], reason: "Navigate to the app and verify the onboarding flow")The response tells you the tier granted for each app. From that point, Claude can call open_application, screenshot, and the interaction tools that the tier permits. The allowlist resets when the server process exits.
The server wraps two native NAPI binaries bundled inside Claude.app:
captureExcluding, captureRegion), display enumeration, app management, TCC permission checks. Its async methods require the macOS run loop to be drained explicitly; the server handles this internally.moveMouse, mouseButton, mouseScroll), keyboard (keys, typeText), cursor position, frontmost app info.No Electron, no AppleScript, no Accessibility API polling. Both binaries dispatch real CGEvents directly into the macOS event system.
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.