DataClaw
A one-command CLI scrapes the full agentic session logs off your own disk, every tool call, token count, and model reply from Claude Code, Codex or Gemini CLI. It redacts secrets, then pools the lot into a shared Hugging Face tag anyone can train on.
Run `dataclaw export` and it reads the raw session files your coding agent leaves on disk — not the tidy chat transcript, but the whole trajectory: your prompts, the model's thinking, every tool call it fired, the token accounting. It strips out API keys, private keys and emails, makes you review the result, then uploads it to Hugging Face under a single tag, `dataclaw`. Every user who does this adds to the same pile. Roughly three dozen such datasets already exist, from a dozen different people, one of them a Hugging Face staffer.
A tool built as the responsible way to share coding data already has an unredacted fork in the pool.
Open conversation corpora aren't new — ShareGPT, WildChat and others scraped chat URLs people had already chosen to share. DataClaw's shift is what it takes and how it coordinates: the local agentic run rather than the prose, and a tag-by-convention rather than a curated release, so the commons assembles itself as strangers opt in. It is also, openly, a political act. The README accuses Anthropic of training on the open web and then locking its own data away — 'pulling up the ladder behind them' — and pitches the tool as throwing that ladder back down, a jab at the data-lockdown posture the frontier labs adopted after this year's distillation fights.
The safety story is the soft spot, and worth reading before you run it. Redaction is voluntary, the review gate is skippable, and the README says plainly it is 'NOT foolproof.' The proof is already in the pool: among the community datasets sits one named `Claude-Opus-Dataclaw-Unredacted`. A tool built as the responsible way to donate coding data has, within days, sprouted the irresponsible fork — which is the honest measure of how a self-assembling data commons actually behaves once it leaves the author's hands.
`pip install -U dataclaw`, then `dataclaw export --no-push` to inspect exactly what it pulled off your disk before anything leaves your machine.
Open the repo at github.com →The lenses
The facts
How this connects
Tap a node to open it