PaperClaw

CLI tool that ingests unorganized PDF documents from an inbox folder, classifies each one using Claude AI, generates a human-readable filename, extracts a plain-text transcript, and files everything into a structured library.

• Classifies and renames PDFs using Claude Sonnet
• Extracts full-text transcript as a .md file alongside each PDF
• Organizes library by category and year
• Maintains a machine-readable index.json of all processed documents
• Detects drift between index and filesystem
• Archives originals before any operation

• Python 3.12+
• Anthropic API key

pip install -e ".[dev]"

Copy .env.example to .env and set your API key:

cp .env.example .env
# edit .env and set ANTHROPIC_API_KEY

Process inbox:

paperclaw process

Check for drift between index and filesystem:

paperclaw check

Review last N classifications:

paperclaw review --last 10

~/inbox/                              ← drop PDFs here
~/archive/                            ← immutable originals
~/library/
  index.json
  invoices/2024/
  bills/2025/
  contracts/
  bank-statements/
  tax/
  other/

Files are named {YYYY-MM-DD}_{vendor}_{short-description}.pdf with a matching .md transcript.

Documents with classifier confidence < 0.8 are filed under other and flagged for manual review.

# Run all checks
pre-commit run --all-files

# Run tests
pytest

# Type check
mypy --strict paperclaw/

All tool configuration lives in pyproject.toml.

• Only first-page text is sent to the Claude API (account numbers and transactions on later pages are never transmitted)
• ANTHROPIC_API_KEY is read from .env, which is gitignored
• Originals are archived before any move or rename — zero data loss guarantee
• index.json is updated atomically to prevent corruption