paperloom 0.1.0

Local-first document toolkit. Streaming OCR (GLM-OCR via Ollama), PII anonymizer, and 19 chainable PDF/Markdown/HTML tools. MCP server included.

paperloom

Local-first document toolkit. Streaming OCR + PII anonymizer + 19 chainable tools. MCP-native.

paperloom is the Python library and MCP server behind paperloom — a local-first web app for OCR, PDF/Markdown/HTML transforms, and PII redaction. Every tool runs on your machine. No cloud round-trips, no telemetry.

Why paperloom

paperloom rides a state-of-the-art OCR model — GLM-OCR scores 94.62 on OmniDocBench V1.5 (rank #1) and is SOTA on formula / table recognition and information extraction. paperloom commits to tracking the current SOTA: when a stronger open model ships, the Ollama pin gets updated.

paperloom's value-add is agent orchestration around the model:

• 19 chainable tools — pdf-to-images → ocr → anonymize → markdown-to-pdf in one call.
• MCP server with security model — register_file + path allowlist + file_id tokens. Drop-in for Claude Desktop, Claude Code, Cursor, Cline, Agno.
• Built-in PII redaction — OPF model, 8 entity categories, verbatim.
• Streaming SSE — Markdown emits page-by-page as the OCR model writes it.
• One Ollama dep — reuses any GLM-OCR model you already pulled. No multi-GB model zoo download.

For raw model quality on dense scientific PDFs, marker, docling, and MinerU are excellent companion projects — paperloom doesn't try to out-research them, it focuses on the orchestration and privacy layer around the model.

Install

# Library + CLI + MCP server (no PDF rendering, no anonymizer):
uvx paperloom doctor

# Full toolkit:
uvx --with 'paperloom[all]' paperloom doctor

# Or pip:
pip install paperloom            # core
pip install 'paperloom[pdf]'     # + WeasyPrint (markdown→pdf, html→pdf)
pip install 'paperloom[all]'     # everything published on PyPI

pdf extra needs native libs (brew install pango on macOS).

The OPF anonymizer is not a PyPI extra — it's distributed as a git repo. The anonymize tool auto-installs it on first call (~250 MB Python deps + ~4 GB checkpoint). To opt out and install manually:

PAPERLOOM_AUTO_INSTALL_OPF=0  # disables the auto-installer
uv pip install 'opf @ git+https://github.com/openai/privacy-filter@main'

Use as a library

from paperloom import ocr_to_markdown, anonymize, Chain

# One-shot OCR
md = ocr_to_markdown("scan.pdf")

# Redact PII
clean = anonymize(md, preset="balanced")

# Compose tools
result = Chain([
    ("pdf-to-images", {"dpi": 200}),
    ("ocr-to-markdown", {}),
    ("anonymize", {"preset": "recall"}),
]).run(["doc.pdf"])

Use as an MCP server

uvx --from paperloom paperloom-mcp

Wire into Claude Desktop:

{
  "mcpServers": {
    "paperloom": {
      "command": "uvx",
      "args": ["--from", "paperloom", "paperloom-mcp"],
      "env": {
        "PAPERLOOM_MCP_ALLOWED_DIRS": "/Users/you/Documents,/Users/you/Downloads"
      }
    }
  }
}

Use from the CLI

paperloom ocr scan.pdf -o out.md
paperloom anonymize out.md --preset recall
paperloom chain --steps pdf-to-images,ocr-to-markdown,anonymize doc.pdf
paperloom doctor      # check Ollama, glm-ocr, OPF, allowlist

Requirements

• Ollama with glm-ocr:latest pulled (ollama pull glm-ocr:latest).
• Python 3.11+.
• Hardware: GLM-OCR is RAM-bound. Recommended: Apple Silicon M-series Pro with ≥ 24 GB unified RAM, or x86 with 24 GB+ GPU VRAM. Minimum 16 GB. Below 16 GB the OS can freeze hard enough to require a reboot — don't.
• (Optional) pango for the pdf extra; ~4 GB checkpoint for the anonymizer extra (auto-downloaded on first call).

License

MIT. Depends on OpenAI Privacy Filter (Apache 2.0) when the anonymizer extra is installed.