Wireshark for AI agents, but it lives in the kernel

What it does

MCPSpy is a Linux CLI tool that hooks kernel-level system calls with eBPF to capture JSON-RPC 2.0 messages flying between MCP clients and servers. It decodes both stdio and HTTP/HTTPS transports into a browsable TUI or structured JSONL logs, effectively turning opaque AI agent chatter into observable events. The tool also offers optional prompt-injection scanning via HuggingFace inference models, flagging potential jailbreaks in tool arguments or responses.

The interesting bit

Most debugging tools sit in userland; MCPSpy plants itself in the kernel using eBPF, which means it can observe inter-process MCP traffic without instrumenting the applications themselves. That architecture makes it viable as a Kubernetes DaemonSet, passively auditing containers running LangFlow, LangGraph, or any other MCP-speaking service from the node level.

Key highlights

• Passively monitors stdio pipes and HTTP/S streams, including Server-Sent Events
• Optional real-time security analysis using external ML models to detect prompt injection
• Outputs rich JSONL with full request/response correlation and raw message capture
• Ships as a single binary with an interactive TUI, or runs headless for log collection
• Supports deployment as a privileged container or Kubernetes DaemonSet for cluster-wide visibility

Caveats

• Linux-only (kernel 5.15+) and requires root privileges, so forget about monitoring from your MacBook directly
• Prompt-injection detection depends on a HuggingFace API token and external inference, which means it is neither offline nor self-contained
• The README notes that one of the supported security models is already deprecated on HuggingFace’s inference API

Verdict

Worth a look if you run MCP workloads on Linux and need to debug opaque client/server conversations or audit them for compliance. Skip it if your stack is Windows, macOS, or you were hoping for a fully offline, batteries-included security scanner.