What it does DataClaw scrapes conversation history from Claude Code, Codex, and other coding agents, packages it into structured JSONL, and publishes it to Hugging Face as a dataset. It ships as both a Python CLI and an unsigned Mac menu-bar app. The author frames it as “performance art” — a response to Anthropic’s data-sharing restrictions.

The interesting bit The privacy workflow is unusually paranoid, in a good way. Before you can push anything, you must run a local export, review it, optionally scan for your full legal name, and sign attestations confirming you checked for PII. The tool also auto-redacts secrets via regex, entropy analysis, and username hashing. It is designed so you can literally paste a six-step prompt into Claude Code and have the agent export itself.

Key highlights

• Parses session logs including voice transcripts, images, tool calls, and token usage
• Multi-layer redaction: API keys, emails, usernames, high-entropy strings, plus custom rules
• dataclaw jsonl-to-yaml and diff-jsonl for human review before publishing
• Tagged dataclaw on Hugging Face to form a distributed open dataset
• Mac app bundles everything; CLI works everywhere Python does

Caveats

• Mac DMG is currently unsigned, so Gatekeeper will complain on first launch
• Intel Macs must use the CLI; Apple Silicon only for the app
• The author explicitly warns automated redaction is “NOT foolproof”

Verdict Worth a look if you believe coding-agent transcripts should be public training data and you are willing to do the privacy homework. Skip it if you just want a backup of your chats — this tool wants you to publish.