Turn any document into structured data.

One-click install for AI agents

docfold ships a CLI (docfold) and an MCP server (docfold-mcp) with four tools — parse_document, extract_tables, list_engines, classify_document. All four definitions cost your agent under 1000 tokens (budget-tested). Failures come back structured ({error, failures}) — never garbage posing as content. One command registers it in your client:

$ pip install "docfold[mcp,pymupdf]"
$ docfold install claude   # Claude Code
$ docfold install codex    # Codex CLI
$ docfold install cursor   # Cursor
$ docfold install vscode   # VS Code

Any other client: paste the config JSON above into its MCP settings. Full agent instructions live at install.md — fetchable by any agent.

Why Docfold?

Every document parser has trade-offs — speed, tables, scans, cost. Docfold lets you switch between them with one line, and falls back automatically when an engine fails on a file.

20+ engines, one interface

Local engines are free and fast; OCR engines handle scans; cloud engines add managed accuracy. Every engine is an optional extra — install only what you use.

How to choose

Or skip the decision entirely — call router.process(file) and let extension-aware routing pick.

• Born-digital PDF (has a text layer)PyMuPDF — sub-second, zero GPU
• Complex layout, tables, reading orderDocling or Marker
• Scanned PDF or photo of a documentTesseract (free) → Surya / Mistral OCR (better)
• Office files (DOCX, XLSX, PPTX)MarkItDown or Unstructured
• Scientific papers with formulasMinerU or Nougat
• Invoices and forms with key-value fieldsAzure Doc Intelligence / AWS Textract
• Not sure which engine is best for your corpusdocfold evaluate — benchmark them on your own files
• Need an MCP server for AI agentsdocfold-mcp — built-in; register with docfold install claude

Quickstart

Install one extra per engine, or docfold[all] for everything.

import asyncio
from docfold import EngineRouter, OutputFormat
from docfold.engines.pymupdf_engine import PyMuPDFEngine
from docfold.engines.docling_engine import DoclingEngine

async def main():
    router = EngineRouter([PyMuPDFEngine(), DoclingEngine()])

    # Auto-routed by extension; falls back if an engine fails
    result = await router.process("report.pdf")
    print(result.content)         # LLM-ready markdown
    print(result.engine_name)     # which engine actually parsed it

    # Force an engine, get JSON with tables and bounding boxes
    result = await router.process(
        "invoice.pdf",
        output_format=OutputFormat.JSON,
        engine_hint="docling",
    )

asyncio.run(main())

# CLI
$ docfold convert report.pdf                 # auto-routed → markdown
$ docfold convert scan.pdf -e tesseract -f json
$ docfold compare report.pdf -e pymupdf,docling  # side-by-side quality check
$ docfold engines
$ docfold doctor            # health check: engines, MCP extra
$ docfold update --check    # self-update via PyPI

Built by & ecosystem

Docfold is built and maintained by Mike Sadofyev (CEO, Datatera.ai) — the document engine behind Datatera — alongside a small ecosystem of AI-data tooling. Connect on LinkedIn, X, or GitHub.