Azure Document Intelligence MCP connector for Power Platform

April 13, 2026

Microsoft’s MarkItDown is a popular open-source Python utility for converting documents to Markdown. With 106k GitHub stars and support for PDFs, Office files, images, and audio, it’s a go-to tool for developers feeding documents into LLM pipelines. It even has an MCP server for Claude Desktop integration.

But MarkItDown is a developer tool. It requires Python, runs locally or in a container, and has no built-in authentication, compliance controls, or enterprise deployment model. If you need document intelligence inside Power Platform—where a Copilot Studio agent extracts invoice fields, a Power Automate flow classifies incoming documents, or a Power App reads receipts—you need something that speaks the platform’s language.

This connector wraps the Azure Document Intelligence REST API (v4.0 GA, 2024-11-30) with 20 operations for Power Automate and Power Apps, plus 11 MCP tools for Copilot Studio agents. Same AI models that power Microsoft 365 Copilot’s document understanding, deployed behind your Azure subscription with your compliance boundaries.

Full source: GitHub repository

MarkItDown vs. Document Intelligence connector

MarkItDown excels at bulk Markdown conversion for LLM training pipelines. This connector excels when you need structured extraction—pulling specific fields from invoices, classifying document types, or building custom models for your domain-specific forms.

Supported prebuilt models

11 MCP tools

When added as an MCP connector in Copilot Studio, the agent gets 11 curated tools:

20 REST operations

For Power Automate and Power Apps, 20 operations cover the full API surface:

Document models (10 operations)

• Analyze document (async—returns Operation-Location header)
• Get analyze result
• Analyze document output (produces searchable PDF)
• List models
• Get model
• Delete model
• Build custom model
• Compose model
• Authorize model copy
• Copy model

Document classifiers (6 operations)

• Build classifier
• List classifiers
• Get classifier
• Delete classifier
• Classify document
• Get classify result

Operations and service (3 operations)

• List operations
• Get operation
• Get service info

How it works

Document analysis is async. The POST request returns HTTP 202 with an Operation-Location header. Poll until the status is succeeded.

User: "Extract the line items from this invoice"

1. Agent calls analyze_document({
     model: "prebuilt-invoice",
     document_url: "https://storage.blob.core.windows.net/invoices/april-2026.pdf"
   })

   → Returns operation ID: "3b9a21c4-..."

2. Agent calls get_analyze_result({
     operation_id: "3b9a21c4-..."
   })

   → Returns structured JSON:
     Vendor: "Zava Corp"
     Invoice Date: "2026-04-01"
     Total: "$12,450.00"
     Line Items:
       - "Cloud hosting - $8,200.00"
       - "Support plan - $4,250.00"

The MCP script handles polling automatically for Copilot Studio agents. In Power Automate flows, use the separate POST + GET operations with a Do Until loop.

Handling async operations in Power Automate

┌─────────────────────────────┐
│ Analyze Document (POST)     │
│ Model: prebuilt-invoice     │
│ Document: invoice.pdf       │
└─────────────┬───────────────┘
              │ HTTP 202 + Operation-Location
              ▼
┌─────────────────────────────┐
│ Do Until: status = succeeded│
│ ┌─────────────────────────┐ │
│ │ Get Analyze Result      │ │
│ │ (GET Operation-Location)│ │
│ └─────────────────────────┘ │
│ Delay: 2 seconds            │
└─────────────┬───────────────┘
              │ status: succeeded
              ▼
┌─────────────────────────────┐
│ Parse extracted fields      │
│ Use in downstream actions   │
└─────────────────────────────┘

Use cases

Accounts payable automation: A Power Automate flow monitors a SharePoint library for new invoices. The connector extracts vendor, amount, and line items using prebuilt-invoice. Parsed data flows into Dataverse for approval routing—no manual data entry.

Document triage agent: A Copilot Studio agent accepts uploaded documents, classifies them (invoice, receipt, contract, or ID), and routes each to the appropriate extraction model. The agent returns a structured summary and files it in the right SharePoint folder.

Tax season processing: Use prebuilt-tax.us.w2 and prebuilt-tax.us to bulk-extract fields from employee tax documents. A Power Automate flow processes a folder of scanned W-2s and populates a Dataverse table for payroll review.

Custom model for domain forms: Build a custom extraction model trained on your organization’s specific forms—inspection reports, intake forms, or work orders. Deploy it through the connector and let agents or flows extract your custom fields.

Contract review assistant: A Copilot Studio agent uses prebuilt-contract to extract key terms, parties, and dates from uploaded contracts. The agent summarizes obligations and flags renewal dates, then creates follow-up tasks in Planner.

Authentication

Option 1: API key

1. Go to your Document Intelligence resource in the Azure portal.
2. Navigate to Keys and Endpoint.
3. Copy the resource name (for example, my-doc-intel from https://my-doc-intel.cognitiveservices.azure.com).
4. Copy either Key 1 or Key 2.

Option 2: Microsoft Entra ID (OAuth)

1. Register an application in Microsoft Entra ID.
2. Grant it the Cognitive Services User role on your Document Intelligence resource.
3. Create a client secret.
4. Provide the Client ID, Client Secret, and Resource Name when creating the connection.

Prerequisites

1. An Azure subscription
2. An Azure Document Intelligence resource (or Azure AI Services multi-service resource)
3. An API key or Entra ID app registration with Cognitive Services User role

Known limitations

• Free tier (F0) is limited to 500 pages/month.
• Custom neural model builds have per-resource quotas (check via Get Service Info).
• Maximum document size varies by model (typically 500 MB for prebuilt models).
• Some add-on features (high-resolution OCR, formula extraction) incur additional costs.

Setting up the connector

1. Download the connector files from the GitHub repository
2. Import into Power Platform:

pac auth create --environment "https://yourorg.crm.dynamics.com"
pac connector create --api-definition apiDefinition.swagger.json --api-properties apiProperties.json --script script.csx

Or import manually through Power Automate > Data > Custom Connectors > Import an OpenAPI file.

Resources

• Azure Document Intelligence documentation
• Azure Document Intelligence REST API
• MarkItDown on GitHub
• Full connector source