FankChen/tracecrate

Local-first AI agent trace workbench. Inspect Claude Code, Codex and OTLP logs, compare runs, and export privacy-conscious reports. No backend or API keys.

What it solves

TraceCrate is a privacy-first, client-side workbench for inspecting AI agent traces. It allows developers to analyze the sequence of events, identify patterns, and compare different agent sessions without needing a backend, API keys, or telemetry, ensuring that sensitive trace data remains local to the browser.

How it works

The tool reads trace files (such as Claude Code, Codex, OTLP JSON, and its own native format) and parses them using Web Workers to create a searchable timeline of events. It processes data entirely in-memory, providing tools for filtering, metric recording, and side-by-side comparison of two sessions to identify differences in tool calls and field labels.

Who it’s for

It is designed for developers and researchers working with AI agents who need to debug and analyze agentic workflows, track token usage, and inspect tool inputs/outputs without compromising data privacy.

Highlights

  • Privacy-First Architecture: No backend, no accounts, and no telemetry; all processing happens client-side.
  • Detailed Trace Inspection: Searchable timelines with filters for event kind, status, and duration.
  • Session Comparison: Side-by-side alignment of two sessions to spot differences in tool-call counts and recorded metrics.
  • Controlled Export: Ability to export reports as JSON or standalone HTML, with options to remove timing metadata, token usage, or free text for safer sharing.
  • Broad Format Support: Compatible with Claude Code, Codex, and OTLP JSON formats.

Related

  • Project
    jianruntech/geo-score

    The ​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​​ ​  ​

  • Project
    colinlikescode/NanoGPT-Speedrun-Winner

    An optimized GPT-2 training implementation that uses memory and weight averaging techniques to set a new speed record for the modded-nanogpt challenge.

  • Project
    0814wdwd/solo_jev

    An input layout optimizer that reorganizes rows and fields in datasets to maximize prefix-cache reuse in LLM-based decision models, significantly increasing throughput.

  • Project
    oleksiijko/pmb

    A local-first memory system for AI coding agents that uses SQLite and MCP to store and recall project decisions, lessons, and constraints without cloud dependencies.