Local AI agent trace viewer

See what your coding agent really sent.

claude-tap captures real API traffic from AI coding agents and turns it into a local trace viewer. Inspect prompts, tool calls, token usage, latency, streaming responses, and request diffs without uploading private traces to a hosted dashboard.

View on GitHub Read the trace viewer guide

uv tool install claude-tap
claude-tap --tap-client codex
claude-tap export --format html trace.jsonl

claude-tap local AI agent trace viewer showing requests, tool calls, and token usage — Open a real agent run, inspect each request, and compare context between turns.

Tool calls Inputs, outputs, parameters, and adjacent diffs.

Token usage Prompt, completion, cache, and latency metadata.

HTML export Portable trace files for review and archiving.

Local-first No hosted dashboard required for private traces.

Built for coding agent traces, not generic logs.

Agent runs are shaped by prompts, tool schemas, retries, streaming chunks, and model-specific request formats. claude-tap keeps those pieces visible so debugging can start from the actual trace.

Inspect full request flow

Review system prompts, conversation history, tool schemas, raw request bodies, and reconstructed responses.

Debug behavior with evidence

Use structured diffs to see which message, prompt, tool, or parameter changed between adjacent turns.

Share a trace without a service

Export a self-contained HTML viewer that teammates can open locally for review, support, or incident notes.

One workflow across agent clients.

Use one local trace viewer for Claude Code traces, Codex traces, OpenAI Responses API traces, Gemini traces, and other coding-agent sessions.

Claude Code Codex CLI Gemini CLI Cursor CLI OpenCode Kimi CLI Pi Hermes Agent Qoder CLI Antigravity CLI CodeBuddy CLI

Structured diff between adjacent AI agent requests — Structured request diffs show how context changes from one turn to the next.

Dark mode local trace viewer for long AI agent review sessions — Light and dark viewer modes support longer debugging and review sessions.

From run to review artifact.

Keep the trace local while you inspect it, then export a static HTML artifact only when you need to share.

Run Launch your agent through claude-tap.

Capture Record real API traffic and streaming responses.

Inspect Search turns, tools, usage, latency, and diffs.

Export Save a standalone HTML trace for review.

Start with a local trace.

Use claude-tap when you need to understand what an AI coding agent actually did before you tune prompts, compare tools, or share a run with another reviewer.

Install from GitHub