MCP (Model Context Protocol) server for DocQA — verify AI extraction results against original documents, plus PDF extraction, OCR, and format conversion.

Catches what your AI missed: hallucinated invoice totals, wrong dates, arithmetic errors, mismatched field values — all without sending documents to any LLM.

Quick Start

npx docqa-mcp

Requires DOCQA_API_KEY environment variable. Get your key at: https://api.agentsconsultants.com

Setup

1. Get an API Key

Request a key at api.agentsconsultants.com. Free trial available.

2. Configure Claude Desktop

Add to ~/Library/Application Support/Claude/claude_desktop_config.json:

{
  "mcpServers": {
    "docqa": {
      "command": "npx",
      "args": ["docqa-mcp"],
      "env": {
        "DOCQA_API_KEY": "your-api-key-here"
      }
    }
  }
}

3. Or install globally

npm install -g docqa-mcp

Then configure with command: "docqa-mcp" and your DOCQA_API_KEY env var.

Tools

`docqa-verify` ⭐ Core Tool

Verify AI-extracted fields against the original document. Detects hallucinations, arithmetic errors, and field mismatches.

Basic tier — Pattern validation + arithmetic checks (no document needed):

{
  "extraction": {
    "invoice_number": "INV-2024-0042",
    "date": "03/15/2024",
    "total": "$1,296.00",
    "subtotal": "$1,200.00",
    "tax": "$96.00"
  }
}

Standard tier — Full cross-extraction with independent OCR/PDF re-parse:

{
  "extraction": { "total": "$1,296.00", "date": "03/15/2024" },
  "document": "<base64-encoded-pdf>",
  "callerTool": "tesseract"
}

Gate mode — Simple pass/fail for agent workflow integration:

{
  "extraction": { ... },
  "mode": "gate"
}

Returns: { "pass": true, "confidence": 0.94, "failReasons": [] }

Pricing:

Basic: $0.05/verification — pattern + arithmetic checks
Standard: $0.50/verification — full cross-extraction

`docqa-sample`

Fetch a sample verification report to understand the output format. Free, no API key needed.

`pdf-extract`

Extract structured text and tables from PDF documents.

Input: file (base64 PDF), optional filename
Output: Extracted text, page count, tables, metadata, text quality assessment
Tip: If textQuality: "low", the PDF is likely scanned — use ocr instead

`ocr`

Extract text from images using Tesseract OCR.

Input: file (base64 image), optional language (default: eng), optional filename
Supports: PNG, JPG, TIFF, BMP, WebP
Output: Recognized text, confidence score, character count

`convert`

Convert documents between formats.

Input formats: pdf, md, json, txt
Output formats: txt, md, json
Input: file (base64), from, to
Output: Converted document (base64) + size stats

Why DocQA?

AI extraction tools (OCR, LLMs, form recognizers) hallucinate. DocQA independently re-extracts and cross-verifies:

Problem	DocQA Catch
"Invoice total: $1,296" (actual: $1,269)	✅ Amount mismatch
"Date: 2024-03-15" (actual: 2024-03-51)	✅ Date format invalid
"Tax: $96, Subtotal: $1,200, Total: $1,394"	✅ Arithmetic mismatch
LLM invents a line item not in the PDF	✅ Unreported field detected

vs. Competitors

Feature	docqa-mcp	PDF Extraction MCP	Advanced OCR MCP	Mistral OCR MCP
Cross-extraction verification	✅	❌	❌	❌
Hallucination detection	✅	❌	❌	❌
Arithmetic validation	✅	❌	❌	❌
Gate mode for pipelines	✅	❌	❌	❌
PDF extraction	✅	✅	❌	✅
OCR	✅	❌	✅	✅
Price per doc	$0.05-0.50	varies	varies	varies