What Every AI Coding Agent Request Does – Inside the Open‑Source claude‑tap Tool

Overview

claude‑tap is a local proxy and trace viewer for AI coding agents. It intercepts all API traffic between a client and the remote model API, records each request and response in JSONL, and provides visual inspection.

Key capabilities

Full request visibility : system prompt, conversation history, tool definitions, streaming chunks and token usage are displayed.

Adjacent request diff : character‑level highlighted diff between consecutive requests, showing added/removed messages and system‑prompt changes.

Token usage analysis : input tokens, output tokens, cache hits and creations per request.

Live viewer : --tap-live streams trace data to a browser via Server‑Sent Events.

Offline archive : on exit a self‑contained HTML file is generated for local review or sharing.

Local data storage : all traces remain on the user’s machine; authentication headers are redacted.

Supported clients

Works with nine AI coding CLIs: Claude Code, Codex CLI, Gemini CLI, Kimi CLI, OpenCode, Pi, Hermes Agent, Cursor CLI, Qoder CLI.

Installation

uv tool install claude-tap

pip install claude-tap

Basic usage

Trace the default client (Claude Code): claude-tap Enable live view: claude-tap --tap-live Switch to another client (examples):

# Codex CLI
claude-tap --tap-client codex

# Gemini CLI
claude-tap --tap-client gemini -- -p "hello"

# Kimi CLI
claude-tap --tap-client kimi

# Cursor CLI
claude-tap --tap-client cursor -- -p --trust --model auto "hello"

Run only the proxy: claude-tap --tap-no-launch --tap-port 8080 Open an offline trace archive:

claude-tap dashboard

Implementation details

Clients that allow a custom base URL (Claude Code, Codex CLI) use reverse‑proxy mode: the client’s endpoint is pointed to the local proxy, which forwards the request to the real API.

Clients that cannot change the endpoint (Gemini CLI, OpenCode, Pi) use forward‑proxy mode: traffic is routed through HTTPS_PROXY to the local proxy, which uses a self‑signed CA to decrypt TLS.

All intercepted traffic is written as JSONL traces. In live mode the traces are pushed to the browser via SSE; on termination the traces are bundled into a self‑contained HTML file.

Typical workflow

Install the tool.

Run claude-tap (or with --tap-client for a specific CLI).

Observe real‑time updates in the browser if --tap-live is used.

After the session ends, open the generated HTML file to review request/response payloads, token counts and diffs.

Repository

https://github.com/liaohch3/claude-tap