TraceCrate — Your agent ran. Read what happened.

A privacy-first, client-side AI trace workbench. Inspect the sequence, question the patterns, share deliberately.

简体中文 · Format guide · Privacy · Contributing

Actual live-site screenshot, visually reviewed; 1440 × 2026 full page. All events and values are synthetic demo data, not benchmark results.

TraceCrate turns files you choose into a searchable timeline, recorded metrics, heuristic diagnostics, and side-by-side comparisons. No backend, no telemetry, no accounts, no API keys. It reads traces; it does not run agents or execute recorded commands.

Open the live demo · v0.1.0 release · Contribute · Source on GitHub. Explore synthetic runs without installing anything, or run locally below.

Try the story in 30 seconds

Open the hosted demo and follow the walkthrough below. Local use requires Node.js ≥22.12 (24 recommended) and npm:

git clone https://github.com/FankChen/tracecrate.git
cd tracecrate
npm ci
npm run dev

For local use, open the address printed by Vite. Installation time is separate from the 30-second walkthrough:

1. 0–10s: The two synthetic pagination runs load automatically. Open a tool event in Timeline; search or filter by event kind.
2. 10–20s: Open Insights, then Compare to inspect recorded differences. “Baseline” and “optimized” are manufactured demo labels, not measured improvements.
3. 20–30s: Choose Export report → Structure only → Show redacted preview, then download JSON or a standalone HTML report. Review the full download before sharing.

No Claude/Codex installation or agent commands are necessary. For your own data, use Choose files or drop a supported file. Claude session files are private: explicitly select only files you are authorized to inspect; TraceCrate never scans agent directories. Never contribute actual histories as fixtures.

What you can do

• Follow the recorded sequence: search, event-kind filters, relative times, tool inputs/outputs, and metadata; 100 events per page.
• Read evidence, not invented certainty: reported tokens, explicit errors, elapsed time when present; missing values stay unknown. No cost estimates.
• Spot patterns: explicit tool errors, long tool durations, large outputs, and identical name/input repetitions. Heuristics do not establish causes or intent.
• Compare two sessions: recorded metric differences (B − A) and tool-call counts. Descriptive, not a controlled benchmark.
• Share with a deliberate boundary: strict structure-only export by default; optional best-effort pattern redaction. HTML reports have no scripts or external assets; native JSON can be reimported.
• Keep the workflow local: file reading/parsing in Web Workers, sessions in memory, clear-session cancellation. Reload discards imported sessions and restores the demo; downloaded files remain on disk.

Format support — subsets, not universal ingestion

See formats, synthetic examples, and public upstream references. Usage.input includes cacheRead and cacheWrite; those are subcounts, not additions. Total reported tokens = input + output.

Limits: 20 MiB UTF-8 per file, 20,000 input records and normalized events, nesting depth 60; up to 5 files per selection and 10 sessions in memory (including demos). Web Worker support required. Imports read whole files, not streaming; exports and analysis still use the UI thread. Detail/preview rendering is capped at 50,000 characters, not the full downloaded report. See architecture.

Privacy is a boundary, not a guarantee

The app has no trace-upload or telemetry path. That does not make every environment or export safe. Browser extensions, a compromised browser/device, and modified hosted code can read data. A static host still receives ordinary request metadata, such as IP address and user agent.

Structure-only removes arbitrary free text, original identifiers/names, inputs, outputs, and model names using a strict allowlist. Event relationships, statuses, token metrics, and timestamps/durations remain—and can still be sensitive. Pattern redaction preserves text and can miss secrets. Neither mode guarantees anonymization; inspect every field of the full download. Clearing memory is not secure erasure. Read privacy and export details and security reporting.

Development & publication status

npm run check runs lint, unit tests, and the production build. npm run test:e2e invokes Playwright; browser binaries must be installed separately. See contribution checks.

Verified release · 2026-09-10: v0.1.0 was created at tested commit 52d9ae9b73f815c264a3eb39f5fc3eedc5cb9715. CI run 34462095959 passed lint, TypeScript/build, 107 unit tests and 36 desktop Chromium / Pixel 7 emulation browser tests. Recorded core/adapters line coverage is 96.58% (not UI coverage).

Live verification: Pages run 34462098830 succeeded in build, deployment, and Playwright smoke against the actual public site: subpath assets, cold offline first import using a synthetic Claude fixture after page load, default structure-only HTML download, and no console/page errors. The screenshot above comes from its downloaded evidence artifact. See the verification record. Local browser downloads remain blocked; Firefox/WebKit and screen-reader certification are not claimed. Launch posts and the interaction video are not published.

• CI workflow — Node 24, checks, Chromium E2E.
• Manual Pages workflow — default branch only; enable Pages → GitHub Actions, then dispatch manually. Publication checklist.
• Open contribution tasks and roadmap · Organic launch plan and draft copy · Changelog.

Original TraceCrate project; not affiliated with or endorsed by Anthropic, OpenAI, or OpenTelemetry. MIT license, copyright 2026 TraceCrate contributors.