Docintel

A Document Intelligence MCP server — extract text and structured fields from business documents, route low-confidence extractions to a human review queue, and search everything you've processed. Built on the Model Context Protocol so any MCP client (Claude Desktop, Claude Code, or your own agent) can drive it.

This project packages the operating pattern I've shipped in production document pipelines: extract → score → route, with explicit confidence thresholds and a human in the loop for exactly the cases automation shouldn't decide alone.

Architecture

flowchart LR
    subgraph client [MCP Client]
        A[Claude Desktop / agent]
    end
    subgraph server [docintel-mcp]
        B[process_document]
        C[Extractor<br/>pypdf + pattern fields]
        D{Confidence router}
        E[(Document store<br/>BM25 search)]
        F[(Review queue<br/>JSONL)]
    end
    A -- MCP over stdio --> B
    B --> C --> D
    D -- ">= accept" --> E
    D -- "review band" --> F
    D -- "below review" --> X[rejected]
    A -- search_documents --> E
    A -- review_queue_pending / review_resolve --> F

Every extracted field carries a confidence score and a source snippet, so a reviewer can confirm or correct a value in seconds without reopening the document. Thresholds are explicit and tunable — the difference between a demo and something an operations team will trust.

Tools exposed

Tool	What it does
process_document(path)	Extract text + fields from a .pdf/.txt/.md, route by confidence, index for search
search_documents(query, top_k)	BM25 keyword search across processed documents
get_document_text(document_id)	Retrieve extracted text
list_documents()	List processed documents
review_queue_pending()	Fields awaiting human review
review_resolve(item_id, corrected_value)	Record the human-confirmed value

Quick start

git clone https://github.com/reshma449/docintel-mcp.git
cd docintel-mcp
pip install -e ".[dev]"

# run the test suite
pytest

# run the server directly (stdio transport)
docintel-mcp

Connect from Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "docintel": {
      "command": "docintel-mcp",
      "env": {
        "DOCINTEL_ACCEPT_THRESHOLD": "0.85",
        "DOCINTEL_REVIEW_THRESHOLD": "0.30"
      }
    }
  }
}

Then ask Claude: "Process examples/sample_invoice.txt and show me anything that needs review."

Configuration

Env var	Default	Meaning
DOCINTEL_ACCEPT_THRESHOLD	0.85	Confidence at or above → auto-accept
DOCINTEL_REVIEW_THRESHOLD	0.30	Confidence at or above (but below accept) → human review
DOCINTEL_REVIEW_QUEUE	.docintel/review_queue.jsonl	Where the review queue persists

Design notes

• Why regex + heuristics instead of an LLM call in the default extractor? Determinism. The pipeline shape (extract → score → route) is what matters; the FieldExtractor protocol in extraction.py is a one-method interface, so swapping in an LLM or cloud OCR extractor is a ~20-line change that doesn't touch routing, storage, or the MCP surface.
• Why JSONL for the review queue? It's inspectable with cat, diffable in git, and importable into a spreadsheet — which is how real review teams actually start before anyone builds them a UI.
• Why BM25 and not embeddings? For short business documents, BM25 is strong, explainable ("it matched these terms"), and dependency-free. An embedding index would slot in behind the same search_documents tool without changing the client contract.
• Scanned PDFs produce an explicit warning instead of a silent empty extraction — silent empties are how bad data reaches dashboards.

Project layout

src/docintel_mcp/
  server.py       # FastMCP server + tool definitions
  extraction.py   # text + field extraction (FieldExtractor protocol)
  confidence.py   # thresholds, routing, review queue
  store.py        # BM25 document store
  models.py       # dataclasses shared across the pipeline
tests/            # unit tests for every module

License

MIT