kerlenton/mcpsnoop

What it solves

mcpsnoop is a "Wireshark for MCP," providing a transparent way to monitor and debug the communication between an AI client (like Cursor or Claude Code) and a Model Context Protocol (MCP) server. Unlike other inspectors that act as their own client, mcpsnoop sits directly in the data path, allowing developers to see exactly what the real client sends and what the server responds, including identifying calls the model failed to make or calls with incorrect arguments.

How it works

It operates as a transparent shim that wraps the MCP server command. It forwards bytes verbatim between the client and server while copying every JSON-RPC frame to a hub. This hub then provides a live terminal user interface (TUI) for real-time monitoring. It supports both stdio-based servers and streamable-HTTP servers (acting as a reverse proxy), and can export data as OpenTelemetry spans for structured observability.

Who it’s for

Developers building or debugging MCP servers and AI agents who need to see the actual wire-level communication to troubleshoot tool calls, latency, and protocol errors.

Highlights

• Transparent Monitoring: Sits in the real data path without requiring changes to the client or server code.
• CI/CD Integration: Includes a GitHub Action to fail builds based on errors, invalid frames, or latency budgets.
• Observability: Exports to OpenTelemetry and provides live Prometheus metrics for tool performance.
• Versatile Deployment: Supports stdio and HTTP servers, with helper commands to easily wrap/unwrap Claude Desktop configurations.
• Session Management: Ability to mock captured sessions, diff two sessions, and review past logs for post-mortems.