Learning to recover
Failed commands, wrong guesses and the fix that followed. This is the signal that is hardest to fake.
Redact on your machine, approve exactly what leaves, and keep 70% of net revenue when AI labs license your traces.
Paste this into Claude Code, Codex, Cursor, Gemini CLI or OpenCode. It installs the CLI and stops before anything is uploaded.
Help me set up selltraces so I can sell my anonymized coding-agent sessions. 1. Read https://selltraces.ai/SKILL.md and follow it. 2. Install the CLI: curl -fsSL https://selltraces.ai/install | bash Windows: irm https://selltraces.ai/install.ps1 | iex 3. Run "selltraces login" and show me the code to approve in my browser. 4. Run "selltraces eligibility" and explain anything I still need to do. 5. Walk me through choosing sources and projects, one question at a time. 6. Run "selltraces upload --dry-run" and summarize what would be sent. Rules: never upload, accept terms, or approve projects for me. Ask me first, every time.
No background process and no editor plugin. The CLI reads these folders only when you run it.
~/.claude/projects/**/*.jsonlFull transcripts with tool calls.~/.codex/sessions/**/*.jsonlRollout files, one per session.state.vscdb (workspace storage)Best effort. Cursor changes its format often.~/.gemini/tmp/**/chats/*.jsonSaved chats per project.~/.local/share/opencode/storage/**Sessions, messages and parts.Paths shown for macOS and Linux. Windows paths are in the sources guide.
The CLI reads your sessions, redacts them locally, and reports counts and findings. Nothing is sent.
selltraces upload --dry-runPick sources and projects, mark anything private, add terms to block. Only approved traces upload.
selltraces uploadWe scrub again, score, and list. When a dataset with your traces sells, 70% of its net revenue goes to contributors.
selltraces statusA dry run is the default first step. It prints counts, exclusions and redaction findings by kind. It never prints your content and never contacts the upload API.
sk-ant-api03-•••• becomes [REDACTED:api_key]
Raw traces are stored in private, encrypted storage. Only reviewed, scrubbed versions are packaged for buyers.
No redactor catches everything.
Pattern matching misses secrets in unusual formats, and it cannot judge whether code or a business detail is confidential. Only approve projects you have the right to share, and mark the rest private. Read the limits.
Labs license datasets, not single traces. When a dataset sells, its net revenue is split across the traces in it, weighted by size and quality. 70% of each trace's portion is yours. We keep 30% to run review, storage, sales and payouts.
Before anything sells, the CLI and dashboard show an estimate. It uses a placeholder model of about $1.20 gross for a long, high quality trace, scaled down for short or low scoring ones.
Many traces never sell
Demand depends on what labs are buying.
Paid after the buyer pays
No advances, no guaranteed minimum.
Unsold means removable
Exclude unsold traces at any time.
selltraces
payout statement, example only
Paid after the buyer's payment clears.
Estimates are not payouts.
Synthetic tasks are clean. Real work is not, and models need to learn from the mess.
Failed commands, wrong guesses and the fix that followed. This is the signal that is hardest to fake.
Benchmarks built from tasks people actually asked agents to do, in real repositories.
How developers chain reads, edits, searches and test runs, and when they stop the agent.
Multi-hour sessions show how plans form, drift and get corrected over hundreds of turns.
Buying data? See the lab pageMore in the docs FAQ.
Maybe not. Your employment agreement, client contracts or open source licenses may restrict it. You are responsible for only approving projects you have the right to share. When in doubt, mark the project private and it is never read.
Only traces from sources and projects you approved, after on-device redaction, as normalized JSONL: your messages, the agent's replies, tool calls and tool results. Raw folder paths are replaced with an alias and a hash. The dry run shows the plan first.
At launch, uploads are open to US residents who are 18 or older and have accepted the contributor terms. Anyone, anywhere, can install the CLI and run the local preview.
We do not know yet, and neither does anyone else. It depends on what labs buy and how your traces score. The estimate in the CLI is a placeholder model, not a quote. Some contributors will earn nothing.
Yes, for unsold traces. Run selltraces exclude or use the dashboard and they leave future datasets. Licenses already granted to a buyer for a past sale cannot be pulled back, which the contributor terms explain.
No. There is no daemon, no file watcher and no editor extension. The CLI does work only when you run a command.
You connect a payout account in the dashboard. Earnings become payable after the buyer's payment clears. We will publish the schedule and minimums in the dashboard before the first sale.
The preview is free, local and takes about a minute. Nothing uploads until you approve the plan.
Or install it yourself
$ curl -fsSL https://selltraces.ai/install | bashInstalls one binary to ~/.selltraces/bin. Read the script first at selltraces.ai/install.