codeclaw 0.4.1

Export your Claude Code and Codex conversations to Hugging Face as structured training data

CodeClaw

CodeClaw exports Claude Code and Codex sessions into privacy-safe training datasets, with gated publish controls, automated sync workflows, and optional MCP memory tooling.

Why CodeClaw

• Turn day-to-day coding sessions into structured, reusable training data.
• Keep privacy controls first-class with redaction and manual review gates.
• Preserve historical problem-solving context through MCP-accessible session memory.

Core Capabilities

• Multi-source ingestion:
  • Claude Code and Codex session discovery and parsing.
  • Experimental adapter routing for Cursor, Windsurf, Aider, Continue.dev, Antigravity, VS Code, Zed, and Xcode beta logs.
• Privacy-aware export:
  • Secret and PII redaction, username anonymization, and project-level exclusions.
  • Layered privacy engine: regex baseline + optional ML NER (codeclaw[pii-ml]).
• Controlled publishing workflow:
  • Local export, user review attestations, confirm gate, then push.
  • Immutable dataset version snapshots + dedupe index on publish.
• Continuous mode:
  • Background watch daemon for incremental sync.
• Memory tooling:
  • MCP server with search, project patterns, trajectory stats, session lookup, graph similarity retrieval, and index refresh.

Install

pip install codeclaw

Optional extras:

pip install "codeclaw[pii-ml]"    # Presidio + spaCy detection layer
pip install "codeclaw[mcp]"       # MCP server runtime
pip install "codeclaw[finetune]"  # Experimental local fine-tune scaffolding

From source:

git clone https://github.com/ychampion/codeclaw.git
cd codeclaw
pip install -e ".[dev]"

Quick Start

# Guided onboarding (HF auth help, repo setup, project scope, MCP, watcher)
codeclaw setup

# Verify environment and connected scope
codeclaw doctor
codeclaw projects --source both
codeclaw stats
codeclaw diff --format json
codeclaw config --encryption status

# Export locally first
codeclaw export --no-push

# Review and confirm
codeclaw confirm \
  --full-name "YOUR FULL NAME" \
  --attest-full-name "Asked for full name and scanned export." \
  --attest-sensitive "Reviewed for company/client/private identifiers." \
  --attest-manual-scan "Manually reviewed representative sessions."

# Publish only after explicit approval
codeclaw export --publish-attestation "User explicitly approved publishing to Hugging Face."

# Optional one-command sharing flow
codeclaw share --publish --publish-attestation "User explicitly approved publishing to Hugging Face."

Commands

Command	Description
codeclaw status	Show current stage and next steps (JSON)
codeclaw prep	Discover projects and auth state
codeclaw setup	Guided onboarding (HF, dataset repo, projects, MCP, watcher)
codeclaw doctor	Verify logs, HF auth, and MCP registration
codeclaw stats	Show session, token, redaction, and export metrics
codeclaw stats --skill	Include trajectory-based growth metrics
codeclaw diff	Preview exactly what would be redacted before confirm
codeclaw projects	Manage connected project scope
codeclaw list	List projects with source, size, and exclusion state
codeclaw config ...	Configure repo, sources, exclusions, and redactions
`codeclaw config --encryption on	off
codeclaw export --no-push	Export locally for review
codeclaw confirm ...	Run checks and unlock push gate
codeclaw export --publish-attestation "..."	Push dataset after approval
codeclaw share [--publish]	Fast export flow with optional publish + dataset card update
`codeclaw watch --start	--stop
codeclaw serve	Start MCP server over stdio
codeclaw install-mcp	Register MCP server in Claude config
codeclaw synthesize --project <name>	Generate CODECLAW.md from synced sessions
codeclaw update-skill claude	Install/update local CodeClaw skill

Experimental preview command:

• codeclaw finetune --experimental

Additional source filters are available for adapter-backed ingestion:

• cursor, windsurf, aider, continue, antigravity, vscode, zed, xcode-beta

MCP Memory Server

Install optional MCP dependency:

pip install "codeclaw[mcp]"
codeclaw install-mcp

Available MCP tools:

• search_past_solutions(query, max_results=5)
• get_project_patterns(project=None)
• get_trajectory_stats()
• get_session(session_id)
• find_similar_sessions(context, max_results=5)
• refresh_index()

Privacy and Safety

CodeClaw is designed for private-by-default workflows:

• path and username anonymization
• secret and high-entropy token detection
• custom redaction lists
• manual confirmation and attestation gates before publish
• encryption-at-rest support for local artifacts with keyring-backed key management

Automated redaction is not perfect. Always review local exports before publishing.

Package Distribution

• Primary: PyPI (pip install codeclaw)
• Additional: GitHub Packages publish workflow is included for org/internal registry consumption.

Install from GitHub Packages:

pip install codeclaw \
  --index-url https://pypi.pkg.github.com/ychampion/simple/ \
  --extra-index-url https://pypi.org/simple

Community

• Contribution guide: CONTRIBUTING.md
• Security policy: SECURITY.md
• Support channels: SUPPORT.md
• Code of conduct: CODE_OF_CONDUCT.md
• Release process: RELEASE.md

License

MIT - see LICENSE.