cc-tracer 0.1.3

A local, infra-free, raw-fidelity inspector for a single Claude Code session — a DevTools Network tab for Claude Code.

cc-tracer (For Claude Code)

A local, infra-free, raw-fidelity inspector for a single Claude Code session — a DevTools Network tab for Claude Code. Install it, point Claude Code at it, and watch one session's hook events and raw API turns on a live timeline.

How it captures

A local FastAPI server (127.0.0.1:7355) with Two capture tiers:

• Hook tier — Claude Code hooks POST to the tracer server. Captures what Claude did: prompts, Pre/PostToolUse, results, stop reasons.
• API-proxy tier — point ANTHROPIC_BASE_URL at the tracer; it streams /v1/* to the real API (plain HTTP in, real HTTPS out — no cert trust needed) and reassembles the streaming response into full API turns: system prompt, message context, reasoning text, token usage — what Claude saw and thought.

Getting started

Prerequisites: python3 (3.9+), pip, and the claude CLI on your PATH.

Install the package, then let one command do everything — configure the Claude Code hooks, start the tracer server, and launch Claude Code routed through it:

pip install cc-tracer          # or: pip install -e .   (from a clone)
cc-tracer start

Then open the tracer UI at http://127.0.0.1:7355 and use Claude Code as usual. The tracer keeps running after you quit Claude Code so you can keep browsing the capture — run cc-tracer stop when you're done.

What cc-tracer start does:

1. Merges the tracer hooks into the project-local .claude/settings.json (in the current directory, so they apply only to this project — not every Claude session; idempotent, backs the original up to .bak once). Use --settings PATH to target a different file, e.g. ~/.claude/settings.json to trace globally.
2. Starts a detached tracer server on :7355 (hook sink + UI + API proxy), or reuses one already running there. The server is left running when Claude exits.
3. Exports ANTHROPIC_BASE_URL=http://127.0.0.1:7355 and runs claude (you can pass claude args after --).

Session JSONL is written to ~/.cc-tracer/logs/ (override with TRACER_LOG_DIR or --log-dir).

For instance:

cc-tracer start \
  --log-dir ./logs/ \
  -- -c # -c to continue recent claude session

To run just the server without launching Claude — e.g. to browse past captures in the UI — use

cc-tracer start --server-only

Stopping

cc-tracer stop    # stop the server AND remove the tracer's hooks

Hooks follow the server's lifecycle: start adds them and stop removes only the tracer's own entries (your other hooks are left alone). With the server running, you can also point a Claude session you launch yourself at it with export ANTHROPIC_BASE_URL=http://127.0.0.1:7355.

When to use this vs. OpenTelemetry

Claude Code emits OpenTelemetry natively, and tools like SigNoz, Grafana, and LangSmith build on it. They are aggregate observability — dashboards, cost/latency trends, alerting, cross-session correlation across a fleet. If that's your goal, use them; this tracer doesn't try to.

This tracer targets the thing those tools explicitly aren't: a local, infra-free, raw-fidelity inspector for a single session — think "DevTools Network tab for Claude Code."

	This tracer	Native OTel (SigNoz / Grafana / LangSmith)
Raw API request/response body	✅	❌ (spans/metrics; content redacted by default)
Reasoning / thinking text	✅	❌
Tool layer (Pre/Post, Edit diffs)	✅	✅
Backend infra required	none (JSONL + one HTML file)	collector + storage + UI
Data stays fully local	✅	configurable / often SaaS
Cross-session stats & alerting	❌	✅

The closest off-the-shelf analog is a manual mitmproxy setup; the base-URL forwarder avoids that approach's CA forging and NODE_EXTRA_CA_CERTS fuss.

Caveats

• Proxy ≠ full coverage. The API proxy only sees HTTP traffic. Transport that isn't plain HTTP (e.g. the Agent SDK's IPC/WebSocket) is captured only at the hook tier. Native OTel emits regardless of transport.
• SSE reassembly is schema-coupled. Rebuilding turns depends on the current Anthropic event shape, so it needs upkeep when the API evolves — a maintenance cost the OTel-based tools don't carry.