Microsoft Copilot Studio Integration
with MCP Server - Overview

Connect your MCP Server to Microsoft Copilot Studio to create low-code/no-code AI agents with enterprise data access. Perfect for business users and citizen developers.

Table of contents

• What You'll Build
• Prerequisites
• Quick Start
• Advanced Features
• Industry Templates
• Sample Copilot Export

What You'll Build

• Custom Copilot with MCP tool integration
• Conversational AI for domain-specific tasks
• Enterprise Chatbot with Azure service access
• Teams/Web deployment ready

Prerequisites

• Microsoft Copilot Studio license
• MCP Server deployed with public endpoint
• Microsoft Teams (optional, for Teams deployment)

Quick Start

Step 1: Access Copilot Studio

1. Navigate to Copilot Studio
2. Sign in with your Microsoft account
3. Select your environment (or create new)

Step 2: Create New Copilot

1. Click Create → New Copilot
2. Choose From Blank
3. Enter details:
4. Name: Healthcare Assistant (or your industry)
5. Description: AI assistant with MCP-powered data access
6. Language: English
7. Click Create

Step 3: Add MCP Server as Tool

Option A (Recommended): MCP onboarding wizard:Copilot Studio supports connecting directly to an MCP server using the Model Context Protocol tool type.

1. In your Copilot, go to the Tools page
2. Select Add a tool → New tool
3. Select Model Context Protocol
4. Enter:
5. Server URL: https://your-mcp.azurecontainerapps.io/mcp
6. Authentication: None, API key, or OAuth 2.0

Note

• Copilot Studio currently supports the Streamable transport type for MCP.
• This repo’s MCP server exposes a standards-based Streamable endpoint at POST /mcp (JSON-RPC).

Option B: Custom connector (MCP Streamable): If you need to manage the connection via Power Apps, create a custom connector using a minimal OpenAPI schema that points to POST /mcp and includes the MCP protocol marker.

swagger: '2.0'
info:
  title: Azure MCP Blueprint
  description: MCP Streamable endpoint for Copilot Studio
  version: 1.0.0
host: your-mcp.azurecontainerapps.io
basePath: /
schemes:
  - https
paths:
  /mcp:
    post:
      summary: MCP Streamable endpoint
      x-ms-agentic-protocol: mcp-streamable-1.0
      operationId: InvokeMCP
      responses:
        '200':
          description: JSON-RPC response
        '202':
          description: Accepted (notifications / responses)

Step 4: Configure Topics with MCP Tools

Example: Healthcare Patient Lookup

1. Go to Topics tab
2. Click + New topic → From blank
3. Name: "Patient Lookup"
4. Add trigger phrases:
5. "Find patient"
6. "Search for patient records"
7. "Show me patient information"

Conversation Flow:

Node 1: Trigger - Patient lookup request detected

Node 2: Question - Get patient details
  - "What is the patient name or ID you're looking for?"
  - Save response as: patientQuery

Node 3: Action - Call MCP Tool
  - Tool: `search_documents`
  - Arguments:
    {
      "query": "{patientQuery}",
      "top": 5
    }
  - Save response as: searchResults

Node 4: Message - Display results
  - "I found the following patient records:"
  - {searchResults.content.results[0].firstName} {searchResults.content.results[0].lastName}
  - Patient ID: {searchResults.content.results[0].patientId}
  - Last Visit: {searchResults.content.results[0].lastVisitDate}

Node 5: Question - Follow-up
  - "Would you like more details on any patient?"
  - Options: Yes / No

1. Save topic
2. Test in Test Copilot pane

Step 5: Industry-Specific Topics

Example topic mappings (Industry → Trigger → MCP tool call)

Industry	Topic	Trigger phrase examples	MCP tool	Example action (tool → parameters/query)
Healthcare	Medication Lookup	“What medications is patient taking?”, “Show meds for {patientId}”	cosmos_query_items	cosmos_query_items → SELECT c.medications FROM c WHERE c.patientId = '{patientId}'
Healthcare	Allergy Check	“Check patient allergies”, “Is {patientId} allergic to {allergy}?”	search_documents	search_documents → {"query":"","top":10,"filter":"allergies/any(a: a eq '{allergy}')"}
Retail	Product Search	“Find product”, “Search {product}”, “Products under $50”	search_documents	search_documents → {"query":"{productQuery}","top":10,"filter":null}
Retail	Inventory Check	“Check stock availability”, “Is SKU {sku} in stock?”	cosmos_query_items	cosmos_query_items → SELECT * FROM c WHERE c.sku = '{sku}'
Finance	Transaction Search	“Show my transactions”, “Transactions last 30 days”	cosmos_query_items	cosmos_query_items → SELECT * FROM c WHERE c.accountId = '{accountId}' ORDER BY c.timestamp DESC
Finance	Fraud Alert	“Check for suspicious activity”, “High fraud scores”	cosmos_query_items	cosmos_query_items → SELECT * FROM c WHERE c.fraudScore > 0.7 ORDER BY c.fraudScore DESC

Step 6: Generative Answers (Optional)

Publish to Demo Website: Enable generative responses powered by MCP data.

1. Go to Settings → Generative AI
2. Enable Generative answers
3. Configure:
4. Data source: MCP Server (via connector)
5. Moderation: Medium

6. Content safety: Enabled

7. Create Generative Topic:

System Instructions:
You are a {industry} AI assistant with access to real-time data via MCP tools.

Available Tools:
- search_documents: Full-text search
- cosmos_query_items: SQL-like queries
- search_semantic: AI-powered semantic search
- openai_chat_completion: Generate insights

When users ask questions:
1. Identify the appropriate MCP tool
2. Call the tool with correct parameters
3. Interpret results clearly
4. Provide helpful, accurate responses

Always maintain data privacy and security.

Step 7: Test Your Copilot

1. Click Test your copilot (top right)
2. Try example queries:
3. "Find all diabetic patients"
4. "Search for products under $50"

5. "Show high-value transactions"

6. Verify MCP tool calls in Test pane

Step 8: Publish

Publish to Demo Website:

1. Go to Publish tab
2. Click Publish
3. Select Demo website

4. Share link: https://your-copilot.powerapps.com/...

   Publish to Microsoft Teams:

5. Go to Publish tab

6. Click Publish
7. Select Microsoft Teams
8. Configure:
9. Icon
10. Short description
11. Full description
12. Submit for approval (if required)

13. Install in Teams

    Embed in Website:

<!DOCTYPE html>
<html>
<head>
    <title>Healthcare Assistant</title>
</head>
<body>
    <h1>Healthcare AI Assistant</h1>

    <!-- Copilot Studio Embed Code -->
    <div id="copilot-container"></div>
    <script src="https://cdn.botframework.com/botframework-webchat/latest/webchat.js"></script>
    <script>
        window.WebChat.renderWebChat({
            directLine: window.WebChat.createDirectLine({
                secret: 'YOUR_DIRECT_LINE_SECRET'
            }),
            userID: 'user-' + Date.now(),
            username: 'User',
            locale: 'en-US',
            styleOptions: {
                botAvatarImage: 'https://your-logo.png',
                botAvatarInitials: 'HA',
                userAvatarImage: '',
                userAvatarInitials: 'You',
                primaryFont: 'Segoe UI, sans-serif'
            }
        }, document.getElementById('copilot-container'));
    </script>
</body>
</html>

Advanced Features

Authentication & Security

Azure AD Authentication:

1. In Settings → Security
2. Enable Authentication
3. Select Azure Active Directory
4. Configure:
5. Tenant ID
6. Client ID

7. Redirect URI

   Row-Level Security:

# In MCP Server, implement user-scoped queries
def get_user_data(user_id: str, query: str):
    # Add user filter to all queries
    scoped_query = f"{query} AND c.ownerId = '{user_id}'"
    return cosmos_client.query(scoped_query)

Analytics & Monitoring

1. Go to Analytics tab
2. View metrics:
3. Total sessions
4. Resolution rate
5. Escalation rate

6. MCP tool usage

7. Export logs for analysis

Multi-Language Support

1. Settings → Languages
2. Add languages:
3. Spanish
4. French
5. German
6. MCP server returns localized results

Industry Templates

Each row maps a user intent to a specific MCP tool call, what to pass, and what to show.

Industry	Topics (deep dive)	Trigger phrases (examples)	MCP tools used	Example calls (shapes)	What to display back	Notes (design + safety)
Healthcare	- Patient Lookup - Medication History - Appointment Scheduling - Lab Results Inquiry - Allergy Checker	- “Find patient”, “Find patients with diabetes” - “Show medications for {patientId}” - “Book with Dr. Smith next week” - “Show latest labs / HbA1c” - “Is {patientId} allergic to penicillin?”	- search_documents - cosmos_query_items - search_semantic - openai_chat_completion	- Search: {"query":"diabetes","top":5,"filter":null} - Cosmos: {"query":"SELECT c.medications FROM c WHERE c.patientId = '{patientId}'"} - Semantic: {"query":"latest lab results HbA1c for {patientId}","top":5} - Foundry: {"messages":[...],"model":"gpt-4o"}	- Patient match list (name, patientId, lastVisitDate) - Medication/allergy fields only (minimal data) - Lab snippets + dates - Clarifying questions for scheduling	- Prefer patientId over names for precision - Keep outputs descriptive (not prescriptive medical advice) - Avoid returning entire patient record unless required
Retail	- Product Search - Inventory Status - Order Tracking - Loyalty Points - Recommendations	- “Search headphones”, “Laptops under $1000” - “Is SKU {sku} in stock?” - “Track order {orderId}” - “My loyalty points” - “Recommend products like {productName}”	- search_documents - cosmos_query_items - openai_chat_completion	- Search: {"query":"headphones","top":10,"filter":null} - Cosmos: {"query":"SELECT * FROM c WHERE c.transactionId = '{orderId}'"} - Foundry: {"messages":[...],"model":"gpt-4o"}	- Product list (name/category/price if indexed) - Availability/stock fields (if present) - Order status + last update - Loyalty point balance - Short recommendation list	- Confirm identifier formats (orderId vs transactionId) - For recommendations: Search first (grounding) then summarize with Foundry - Minimize PII (use customerId, not email)
Finance	- Account Balance - Transaction History - Fraud Alerts - Payment Processing - Financial Insights	- “Balance for account {accountId}” - “Transactions last 30 days” - “Suspicious activity?” - “Pay my bill / send $50” - “Spending insights”	- cosmos_query_items - openai_chat_completion	- Cosmos: {"query":"SELECT * FROM c WHERE c.accountId = '{accountId}' AND c.timestamp >= '{isoDate}' ORDER BY c.timestamp DESC"} - Fraud: {"query":"SELECT * FROM c WHERE c.fraudScore > 0.7 ORDER BY c.fraudScore DESC"} - Foundry: {"messages":[...],"model":"gpt-4o"}	- Balance + currency (if stored) - Recent transactions (amount/merchant/time) - Flagged transactions + fraudScore - Confirmation step for payments - Category insights + next steps	- Present fraud as a signal, not a final determination - Only simulate payments unless you have a real backend - Best pattern for insights: Cosmos query → Foundry summary

Important: Template data is synthetic but can contain PII-like fields such as names, email addresses, phone numbers, addresses, and dates of birth. Avoid logging tool outputs and apply least-privilege access.

Operations area	Symptom / goal	What to check	Fix
Topic design	Topic tries to do too much	One topic is handling multiple intents	Split into multiple topics; keep each topic to a single task.
Fallback	Users ask unhandled questions	No fallback topic / poor trigger phrases	Add a fallback topic; expand trigger phrases; route to a “help me choose” question.
Testing	Tools don’t fire in Test pane	Action node not reached; variables not set	Add explicit questions to capture inputs; verify the tool name and JSON shape; save tool response into a variable.
Monitoring	Need usage + quality visibility	No review cadence	Review Copilot Studio analytics weekly; correlate with MCP server logs (App Insights / platform logs).
MCP connector	Connector fails to connect	Endpoint not public; auth mismatch; CORS; wrong URL	Verify the MCP URL is reachable and points to /mcp; validate auth settings (none/API key/OAuth); configure allowed origins if enforced by the server.
Tool call failures	Tool returns an error	Bad arguments; missing dependent Azure service config	Validate against the tool’s input schema; check MCP server logs; verify Cosmos/Search/Foundry endpoints and credentials/managed identity permissions.

Sample Copilot Export

See the Copilot Studio sample exports in the source repository for:

• Healthcare Assistant (.zip export)
• Retail Assistant (.zip export)
• Finance Assistant (.zip export)