🐕‍🦺 CerbereAG MCP Server

Runtime security for AI agents — as an MCP server.

CerbereAG MCP lets any MCP-compatible AI agent (Claude, Cursor, Cline, custom agents built on the Anthropic API) check its own prompts and tool calls for security risk in real time, using the Model Context Protocol.

It wraps the cerbere-ag SDK's policy engine and exposes it as MCP tools, so an agent can call check_prompt_security or authorize_tool_call the same way it would call any other tool — no separate API integration required.

🚀 Quick start (30 seconds)

1. Install

pip install cerbere-ag-mcp

2. Configure your MCP client

Claude Desktop — add to claude_desktop_config.json:

{
  "mcpServers": {
    "cerbereag": {
      "command": "cerbere-ag-mcp",
      "env": {
        "AGENTGUARD_MCP_COLLECTOR_URL": "https://YOUR_AGENTGUARD_HOST",
        "AGENTGUARD_API_KEY": "ag-your-key",
        "AGENTGUARD_AGENT_ID": "my-agent"
      }
    }
  }
}

Cursor — add to .cursor/mcp.json in your project:

{
  "mcpServers": {
    "cerbereag": {
      "command": "cerbere-ag-mcp",
      "env": {
        "AGENTGUARD_MCP_COLLECTOR_URL": "https://YOUR_AGENTGUARD_HOST",
        "AGENTGUARD_API_KEY": "ag-your-key",
        "AGENTGUARD_AGENT_ID": "my-agent"
      }
    }
  }
}

Restart the client. The tools below become available to the agent automatically — no code changes needed on your side.

🛠️ Available tools

📚 Available resource

⚙️ Configuration

🧩 Running it manually (for testing)

AGENTGUARD_MCP_COLLECTOR_URL=http://localhost:8080 \
AGENTGUARD_API_KEY=ag-your-key \
cerbere-ag-mcp

The server starts on stdio, as expected by MCP clients — it is not meant to be run as a standalone HTTP service.

🔗 Related

• Core SDK (decorator-based integration for Python agents): cerbere-ag
• Dashboard, policy docs, other framework integrations (LangGraph, CrewAI, Composio, HTTP gateway): app.cerbereag.site
• Source: github.com/chrismsmr-celcom/agentguard