cerberus-mcp

MCP (Model Context Protocol) server instrumentation for Cerberus API monitoring.

Drop-in replacement for FastMCP that captures tool, resource, and prompt call metrics and streams them to the Cerberus analytics pipeline.

Installation

pip install cerberus-mcp

Quick Start

Replace FastMCP with CerberusMCP — a single-line change:

from cerberus_mcp import CerberusMCP

mcp = CerberusMCP(
    "my-server",
    cerberus_config={
        "token": "your-api-key",
        "client_id": "your-client-id",
        "ws_url": "wss://your-cerberus-backend:8765",
    }
)

@mcp.tool()
def get_weather(location: str) -> str:
    """Get weather for a location."""
    return f"Sunny in {location}"

@mcp.resource("config://settings")
def get_settings() -> str:
    """Return server settings."""
    return '{"theme": "dark"}'

@mcp.prompt()
def summarize(text: str) -> str:
    """Summarize text."""
    return f"Please summarize: {text}"

All tool calls, resource reads, and prompt invocations are automatically captured with:

• Execution timing (duration_ms)
• Sanitized arguments (sensitive values redacted)
• Error tracking
• Result summaries
• MCP client identity (name, version) and session correlation

Events are sent asynchronously via WebSocket to the Cerberus event_ingest backend using the same pipeline as cerberus-django.

Configuration

Set CERBERUS_DEBUG=true to enable verbose logging.

How It Works

CerberusMCP subclasses FastMCP from the MCP Python SDK and wraps the tool(), resource(), and prompt() decorators. Each handler call is intercepted to capture timing, arguments, results, and errors. Events are queued into a thread-safe queue and sent via a background WebSocket connection — zero impact on your MCP server's response times.

Requirements

• Python >= 3.10
• mcp >= 1.0 (MCP Python SDK)
• websockets >= 12.0
• A running Cerberus backend (cerberus-int)

License

MIT