@tryaidr/sentinel-mcp

Official Model Context Protocol (MCP) Security & Blast Radius Gatekeeper for AI Agents

TRY AIDR Sentinel™ sits between AI Desktop Clients (Claude Desktop, Cursor, Windsurf, Cline) or autonomous agent frameworks and underlying MCP servers. It intercepts tool executions (tools/call), verifies blast radius policies, enforces Human-in-the-Loop (HITL) approvals, and blocks unauthorized or catastrophic actions in real-time.

⚡ Quick Start

Add Sentinel as a wrapper around any standard MCP server in your client configuration:

Cursor (.cursor/mcp.json) / Cline (cline_mcp_settings.json)

{
  "mcpServers": {
    "sentinel-filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@tryaidr/sentinel-mcp",
        "--token", "YOUR_AGENT_TOKEN",
        "--gateway", "https://gateway-thbt.onrender.com",
        "--",
        "npx", "-y", "@modelcontextprotocol/server-filesystem", "."
      ]
    }
  }
}

Claude Desktop (claude_desktop_config.json)

{
  "mcpServers": {
    "sentinel-bash": {
      "command": "npx",
      "args": [
        "-y",
        "@tryaidr/sentinel-mcp",
        "--token", "YOUR_AGENT_TOKEN",
        "--gateway", "https://gateway-thbt.onrender.com",
        "--",
        "npx", "-y", "@modelcontextprotocol/server-terminal"
      ]
    }
  }
}

🛡️ Features

• Prompt Injection & Tool Abuse Shield: Detects when untrusted text hijacks an agent's tool execution.
• Blast Radius Boundaries: Enforces path whitelists and disallows dangerous shell commands (rm -rf, raw device writes, chmod 777).
• Human-in-the-Loop (HITL): Automatically pauses high-risk state-modifying actions until approved by an administrator in the TRY AIDR Dashboard.
• Full JSON-RPC 2.0 Compliance: Seamless drop-in proxy for all MCP tools.

📄 License

Apache-2.0 © TRY AIDR