cursor_admin_mcp

Cross-platform secure MCP server for Cursor agents. Exposes elevated execution tools with mandatory human approval before privilege escalation:

• Windows: run_as_admin — PowerShell with WPF approval dialog and UAC

• Linux: run_as_root — bash with zenity approval dialog and sudo

Requirements

All platforms

• Node.js 20 or later

Windows (run_as_admin)

• Windows 10 or later

• Windows PowerShell 5.1 (default powershell on Windows)

• WPF (PresentationFramework) for the approval dialog

Linux (run_as_root)

• bash

• sudo (user must have sudo privileges)

• zenity for approval and password dialogs

• DISPLAY or WAYLAND_DISPLAY set (GUI session required)

Install example (Debian/Ubuntu):

sudo apt install zenity

Security model

Both tools follow the same pattern: payload isolation, visual approval, privilege escalation, log capture, cleanup.

Windows — run_as_admin

1. The agent sends a PowerShell command string.

2. The server writes the command to an isolated temporary .ps1 file.

3. A wrapper script shows a WPF MessageBox with the exact command payload.

4. If the user clicks Yes, UAC prompts for elevation and the payload runs elevated.

5. If the user clicks No, execution stops and the tool returns Execution denied by user.

6. stdout/stderr are captured to a temporary log file and returned to the agent.

7. All temporary files are deleted after each invocation.

Linux — run_as_root

1. The agent sends a bash command string.

2. The server writes the command to an isolated temporary .sh file.

3. A wrapper script shows a zenity question dialog with the exact command payload.

4. If the user clicks Yes, sudo -A runs the payload using a temporary SUDO_ASKPASS script (zenity --password).

5. If the user clicks No, execution stops and the tool returns Execution denied by user.

6. stdout/stderr are captured to a temporary log file and returned to the agent.

7. All temporary files are deleted after each invocation.

The agent cannot bypass approval or elevation on either platform. Use the tool that matches the current OS.

Build and run

npm install
npm run build
npm start

Other scripts:

• npm run typecheck — TypeScript check without emit

• npm test — run unit tests

Branching and releases

Workflow:

1. Branch from develop for features and fixes.

2. Open pull requests targeting develop. CI runs on every push and PR.

3. When ready to ship, merge develop into main (via PR or fast-forward merge).

4. Each push to main runs the Release workflow: build, test, then create a GitHub release tagged v{version} from package.json.

5. Bump version in package.json before merging to main so each release gets a unique tag. If the tag already exists, the release job skips creation and logs a reminder to bump the version.

To start using develop:

git checkout -b develop
git push -u origin develop

Set develop as the default branch in GitHub repository settings if you want new PRs to target it by default.

Cursor MCP configuration

Add this server in Cursor → Settings → Features → MCP Servers (adjust the path if your clone location differs):

{
  "mcpServers": {
    "cursor-admin-mcp": {
      "command": "node",
      "args": ["C:\\dev\\repos\\others\\cursor_admin_mcp\\dist\\index.js"]
    }
  }
}

On Linux, use the appropriate absolute path to dist/index.js.

After changing MCP settings, reload MCP servers in Cursor.

Manual smoke test

Windows

1. Build the project: npm run build

2. Register the MCP server in Cursor using the config above.

3. Ask the agent to call run_as_admin with: Write-Output "hello"

4. Confirm the WPF dialog shows the exact command.

5. Click Yes, approve UAC, and verify output contains hello.

6. Run again and click No — verify the tool returns Execution denied by user. without a UAC prompt.

7. Confirm no leftover temp files remain under %TEMP%.

Linux

1. Build the project: npm run build

2. Register the MCP server in Cursor.

3. Ask the agent to call run_as_root with: echo "hello"

4. Confirm the zenity dialog shows the exact command.

5. Click Yes, enter your sudo password in the zenity prompt, and verify output contains hello.

6. Run again and click No — verify the tool returns Execution denied by user. without a sudo prompt.

7. Confirm no leftover temp files remain under /tmp.

Calling run_as_admin on Linux or run_as_root on Windows returns a clear platform error without elevation.

SDK note

This project uses @modelcontextprotocol/server@2.0.0-alpha.2 (MCP TypeScript SDK v2 alpha). The v2 API may change before stable release.

License

MIT — see LICENSE.