A lightweight macOS menu bar app that watches your AI coding assistant and blocks dangerous actions before they happen.

一款轻量的 macOS 菜单栏应用，实时监控 AI 编程助手的操作，在危险行为发生前将其拦截。

AI coding tools like Claude Code can run shell commands, delete files, and modify system configs on your behalf. Most of the time that's fine — but occasionally an AI makes a mistake, misunderstands your intent, or gets manipulated by malicious content in your codebase.

AI Guard sits in your menu bar and acts as a last line of defense:

• 🚨 Auto-blocks critical actions (e.g. rm -rf, writing to sensitive paths)
• ⚠️ Asks you to confirm suspicious-but-not-catastrophic actions
• ✅ Silently allows safe actions — no interruptions for normal work
• 📋 Logs everything so you can review what your AI has been up to

AI Guard runs a local HTTP server on 127.0.0.1:47821. AI tools that support hook-based auditing (such as Claude Code via PreToolUse hooks) send each tool call to this server before executing it. AI Guard evaluates the risk and responds with { allow: true } or { allow: false }.

AI Tool → PreToolUse hook → AI Guard (local server) → allow / block

The app never sends your data anywhere. Everything runs locally.

git clone https://github.com/Mel0day/ai-guard.git
cd ai-guard
npm install
npm start

Add a PreToolUse hook to your Claude Code settings (~/.claude/settings.json):

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "",
        "hooks": [
          {
            "type": "command",
            "command": "curl -s -X POST http://127.0.0.1:47821 -H 'Content-Type: application/json' -d '{\"risk\": \"'$(echo $CLAUDE_TOOL_NAME | grep -qiE 'bash|exec' && echo WARN || echo INFO)'\", \"aiDecision\": \"ALLOW\", \"toolName\": \"'$CLAUDE_TOOL_NAME'\", \"reason\": \"Tool use request\"}'"
          }
        ]
      }
    ]
  }
}

For a more sophisticated hook that actually analyzes the tool input and assigns risk levels, see the ai-auditor hook (coming soon).

Launch AI Guard. It will auto-detect Claude Code if you've configured the hook. The shield icon in your menu bar shows your protection status at a glance.

AI Guard accepts POST / with a JSON body:

{
  "risk": "CRITICAL" | "WARN" | "INFO",
  "aiDecision": "BLOCK" | "WARN" | "ALLOW",
  "explanation": "Human-readable reason shown in UI and notifications",
  "toolName": "Bash",
  "params": { "command": "rm -rf /tmp/foo" }
}

Response:

{ "allow": true }
{ "allow": false }

Decision logic:

• risk === CRITICAL or aiDecision === BLOCK → blocked silently, notification shown
• aiDecision === WARN → native confirm dialog, user decides
• Everything else → allowed, logged silently

• macOS 12+
• Node.js 18+
• Electron (installed via npm install)

• Electron — native macOS app, menu bar tray
• Vanilla JS / HTML / CSS — no framework, no bundler
• Built-in PNG generation — shield icon drawn in code, no image assets needed
• ~/.aisec/config.json — persists settings and event history locally

MIT