claude-tap

claude-tap sits between your coding agent and the API and shows you every call it makes. Point it at Claude Code, Codex CLI, Gemini CLI, Cursor CLI, OpenCode, Kimi, Grok, and more, and it captures and displays the actual request and response traffic so you can see exactly what your agent is sending and getting back. Traces stay local, and it auto-redacts the common auth headers.

For anyone debugging why an agent went sideways, this is the difference between guessing and knowing. You stop reasoning about what the agent "probably" did and just read the wire.

Free under MIT, install it with uv or pip. Solo developers and teams both get the same thing: full visibility into agent traffic at no cost.

The catch is what it is under the hood: a man-in-the-middle proxy on your own agent traffic. That's exactly what makes it useful, but it also means your agent's API calls route through it, so trust and the redaction behavior matter. On your own machine debugging your own agents, that's fine. Just understand what you're pointing it at before you run it against anything sensitive.