Real coding-agent sessions, with consent on file.

Licensed datasets from developers who chose to contribute. Redacted on their machines, scrubbed again by us, and documented trace by trace.

Claude Code
Codex CLI
Cursor
Gemini CLI
OpenCode

trace.jsonlsample, abbreviated

{"t":"header","v":1,"source":"claude_code","projectLabel":"payments-api","models":["claude-sonnet-4-5"],"cli":"selltraces/0.4.1"}
{"t":"msg","role":"user","ts":1760000000,"text":"the webhook test is flaky, find out why"}
{"t":"tool_call","name":"Bash","input":{"command":"pnpm test webhooks --runInBand"}}
{"t":"tool_result","name":"Bash","output":"FAIL ... expected 200, received 401","isError":true}
{"t":"msg","role":"assistant","ts":1760000042,"text":"The signing secret is read before the env loads..."}
{"t":"tool_call","name":"Edit","input":{"file_path":"src/webhooks/verify.ts","new_string":"const secret = env.STRIPE_SECRET ?? \"[REDACTED:api_key]\""}}

What comes with a dataset.

Normalized JSONL: Every source mapped to one schema: messages, tool calls, tool results with error flags, and session metadata.
Per-trace provenance: Source agent, models used, turn and tool-call counts, time range, client and server redaction counts, quality score.
Consent on file: Each contributor accepted versioned contributor terms that cover licensing to AI developers. We can show which version.
Exclusions honored: Contributors can pull unsold traces at any time. Datasets are cut from what is listed on the day you license.

Licensing, plainly.

Non-exclusive licenses for training, fine-tuning, evaluation and research.
No attempts to re-identify contributors, and no resale of raw traces.
Most of what you pay goes to the contributors whose traces are in your dataset.
Samples available under NDA before any purchase.

Prefer email? hi@selltraces.ai