Data Export

Export every trace in a namespace for offline analysis, fine-tuning datasets, or archival. Available on personal namespaces and organizations.

Create an export

1. Open Settings → Exports in your namespace.
2. (Optional) Set a preset or custom date range, and/or a visibility filter.
3. Click New Export.

Exports run in the background. The page updates as status changes — no need to refresh.

Download

When the export is ready, use the Download button on the row. Files are named traces-export-{unixtime}.zip, where unixtime is the Unix timestamp at creation.

What's inside

The ZIP contains a manifest and gzip-compressed JSONL files, split into numbered chunks for large namespaces:

• manifest.json — export metadata and record counts
• traces/traces-000001.jsonl.gz — one trace per line: title, agent, model, visibility, created/updated/published times, project and git details, the related trace (for subagents), and token usage and cost when known
• messages/messages-000001.jsonl.gz — one message per line, with its timestamp and token counts
• parts/parts-000001.jsonl.gz — one part per line (tool calls, thinking blocks, etc.)
• analysis/analysis-000001.jsonl.gz — the AI-generated title and analysis of each analyzed trace

Rows link by traceId (the trace's URL id) and messageId. Part content is in contentJson, a JSON document stored as a string.

JSONL loads directly in tools like DuckDB, pandas, and BigQuery. In DuckDB, read_ndjson_auto('parts/*.jsonl.gz') reads every chunk at once.

The export also holds the rest of the namespace's data as JSON files:

• workspace/ — namespace settings, members and their roles, invites, claimed domains, agents, API keys (name and prefix), connected integrations, and surfaces with a copy of each version's HTML
• account/ — only when you export your own personal namespace: your profile, sign-in methods, sessions, namespace memberships, emails sent to you, and your conversations with the Traces agent

Credentials are never exported: no API keys or session tokens, OAuth tokens, invite codes, or token hashes.

Large namespaces take a while to export. The row shows progress as traces are written.

Retention

Exports stay in your history until you delete them.

Delete an export

Use the Delete button on any export row. Pending and running exports can also be deleted; if processing is still running, any generated file is discarded.

Push to HuggingFace

Completed exports can be pushed directly to a HuggingFace dataset repo.

Connect HuggingFace first from Settings → Integrations, and create the destination dataset repo on HuggingFace if you don't have one. The upload button is disabled until the integration is connected.

To upload, click Upload to HuggingFace on a completed export row, then pick the destination repo from a searchable list of your most recently updated datasets across your personal account and any approved HuggingFace orgs. If you don't see it right away, start typing the dataset name to search. You can also set an optional subdirectory path. The trace files (not workspace/ or account/) are pushed as a single commit, and a View on HuggingFace link appears with the commit URL when done.

Limits

Each file in the export must be under 50 MB, and an export pushed in one go must be under 300 MB in total. If an export exceeds this, narrow it with a date range filter.