TraceCrate reads Claude Code, Codex, and OTLP session logs in the browser and compares two runs side by side, with no backend or API keys

FankChen has published TraceCrate v0.1.0, an MIT-licensed workbench that reads AI agent session traces in a browser, and the repository has picked up 96 stars in a day since the release on 10 September. It runs entirely on the client, with no backend, telemetry, or API keys, and reads four formats out of the box.

What it reads

The four inputs are Claude Code's message JSONL, Codex's rollout JSONL with session_meta, turn_context, response_item, and selected event_msg records, OTLP JSON with resourceSpans → scopeSpans → spans and the GenAI attributes conventions, and its own native schemaVersion: 1 JSON report format. That last format is the one the export produces, so a shared report can be reloaded by anyone with the workbench.

What you do with it

Once a trace is loaded, the workbench draws an event timeline with search and filters, at 100 events per page. Tool calls open a detail view with inputs, outputs, metadata, and elapsed time. Heuristics highlight the patterns worth looking at first: tool errors, long-duration steps, unusually large outputs, and repeated calls. Two sessions can be loaded together and compared, showing the metric difference (B minus A) and the tool-call counts side by side.

What comes out

Export produces JSON or standalone HTML, with two redaction modes. Structure-only export removes free text, original identifiers and names, inputs, outputs, and model names through a strict allowlist, keeping timestamps, event relationships, and token counts. Pattern redaction keeps the text and tries to strip secrets by pattern, and the README says out loud that this can miss things. Neither mode is anonymisation; both are boundaries for how much of a session you share when you send a report to a teammate or attach it to a bug.

The tool sits in the gap that grew as coding agents moved from single-shot prompts into longer sessions with subagents and dozens of tool calls. The raw JSONL that Claude Code writes into ~/.claude/projects/ and the rollouts Codex writes are correct records but nobody wants to read them. Anthropic's OTLP-based tracing is now a standard shape for the same data, but reading spans in Jaeger or a general observability tool loses the "conversation" the agent was having. A workbench that reads all three, compares runs, and exports a bounded HTML report puts the debugging step in front of both engineers who write agents and reviewers who did not run the session.

TraceCrate is early: v0.1.0, one push, a live demo at fankchen.github.io/tracecrate/ with a synthetic pagination example, and Node.js 22.12 or later to build. A team already running agents at scale probably has their own trace tooling. Anyone who does not, and is doing everything through the Claude Code and Codex CLIs, has a viewer built for those exact files, with the redaction shipped rather than promised.

Source

Primary source: FankChen/tracecrate on GitHub.