Power MCP: run MCP inside a Power Platform custom connector

January 20, 2026

TL;DR

• Power MCP is an MCP server implemented inside the custom code of a Power Platform connector
• One /mcp endpoint, no external hosting, full MCP handshake, tools/list and tools/call
• Works with Copilot Studio actions; respects OBO tokens and Zero Trust
• Template includes full MCP support (2025-11-25) for tools, resources, prompts, logging
• Built-in tools: echo, get_data; helper methods; optional App Insights telemetry

What is Power MCP?

Power MCP is my implementation of the Model Context Protocol entirely inside a Power Platform custom connector. The connector exposes a single /mcp operation, and the custom C# script handles the MCP handshake, tools/list, and tools/call on the fly. No Azure Function, no separate server—just the connector you already deploy.

Why it matters

• Zero hosting: everything runs in the connector’s custom code
• Single endpoint: /mcp with x-ms-agentic-protocol: mcp-streamable-1.0
• Tool indirection: registry-based dispatch with validation, logging, and schema hints
• Least privilege: OBO token per user; permissions enforced by the underlying service
• Composable: add your domain tools (echo/get_data scaffold included)

Template highlights

Files in the template

📦 Repo: Power MCP Template

Architecture

┌──────────────────────────────────────────────────────────────┐
│                    Copilot Studio Agent                      │
└──────────────────────────────────────────────────────────────┘
                               │ MCP JSON-RPC
                               ▼
┌──────────────────────────────────────────────────────────────┐
│            Power Platform Custom Connector (MCP)             │
│  • /mcp (OpenAPI)                                           │
│  • script.csx (MCP server: initialize, tools/list, tools/call)│
│  • Tool registry (dictionary dispatch)                      │
└──────────────────────────────────────────────────────────────┘
                               │
           ┌───────────────────┴───────────────────┐
           ▼                                       ▼
      Custom APIs                           Microsoft Graph
                                            (or anything else)

Sequence diagram

sequenceDiagram
    participant Agent as Copilot Studio
    participant Connector as Power MCP Connector
    participant Backend as Your API (optional)

    Agent->>Connector: initialize
    Connector-->>Agent: capabilities + serverInfo
    Agent->>Connector: tools/list
    Connector-->>Agent: tools array
    Agent->>Connector: tools/call (echo/get_data)
    Connector->>Backend: GET /data/{id}
    Backend-->>Connector: data
    Connector-->>Agent: tool result
    Agent->>Connector: ping
    Connector-->>Agent: {}

Flow diagram

flowchart TD
    A[Copilot Studio Agent] --> B[/mcp POST/]
    B --> C[InvokeMCP]
    C --> D[script.csx]
    D --> E[Return capabilities]
    D --> F[Build tools]
    D --> G{Execute tool}
    G --> H[Execute echo]
    G --> I[Execute get_data]
    G --> J[Execute get_ticket]
    I --> K[Send external request]
    H --> L[Create JSON-RPC response]
    K --> L
    L --> M[Return to Agent]

Supported MCP methods

OpenAPI: mark the /mcp endpoint as MCP

paths:
  /:
    post:
      operationId: InvokeMCP
      x-ms-agentic-protocol: mcp-streamable-1.0
      summary: MCP endpoint
      description: Handles MCP initialize, tools/list, tools/call
      responses:
        '200':
          description: MCP response stream

MCP JSON-RPC examples

tools/list

{
  "jsonrpc": "2.0",
  "id": 1,
  "method": "tools/list"
}

tools/call → echo

{
  "jsonrpc": "2.0",
  "id": 2,
  "method": "tools/call",
  "params": {
        "name": "echo",
    "arguments": {
        "message": "Hello from MCP"
    }
  }
}

Setup steps

1. Import connector from SharingIsCaring > Power MCP Template
2. Use provided swagger (apiDefinition.swagger.json, path /, basePath /mcp, x-ms-agentic-protocol: mcp-streamable-1.0)
3. Use provided apiProperties to route to custom code (InvokeMCP)
4. Paste script.csx into custom code
5. Configure OAuth for your backend (Dataverse, Graph, custom API)
6. Create a connection and add the connector as an action in Copilot Studio

Testing

# initialize
curl -X POST https://your-connector/mcp -H "Content-Type: application/json" \
    -d '{"jsonrpc":"2.0","method":"initialize","id":1,"params":{"protocolVersion":"2025-11-25","clientInfo":{"name":"test"}}}'

# list tools
curl -X POST https://your-connector/mcp -H "Content-Type: application/json" \
    -d '{"jsonrpc":"2.0","method":"tools/list","id":2}'

# call echo
curl -X POST https://your-connector/mcp -H "Content-Type: application/json" \
    -d '{"jsonrpc":"2.0","method":"tools/call","id":3,"params":{"name":"echo","arguments":{"message":"Hello"}}}'

Copilot Studio: add connector as action → tools surface automatically → test with natural language.

Error handling and stateless behavior

• Protocol errors (-32700, -32601, -32602, -32603) for malformed/unknown methods
• Tool errors returned as successful tool result objects so the model can reason
• listChanged=false; no server-side cache; safe in multi-tenant scenarios

Security

• Least privilege: connector forwards the user’s Authorization header (OBO), honoring Entra permissions
• API keys/tokens: use connection parameters (masked if secret); do not hardcode
• Headers: the connector forwards Authorization
• Zero trust: validate required args (RequireArgument) and fail fast on unknown tools

Limitations

• Static tools: no discovery; tools defined in BuildToolsList
• No streaming: tools/call returns full result; no partial/progress notifications
• No notifications/resources/prompts content: stubs return empty arrays
• Payload size: large responses may hit connector limits—paginate and filter ($top, server-side paging)
• Timeouts/retries: no retries/backoff; long-running calls may time out
• JSON only: responses parsed as JSON; non-JSON APIs must be transformed

Resources

• Advanced (coming soon): Orchestration patterns post
• GitHub: Connector-Code/Power MCP Template
• MCP: modelcontextprotocol.io

#PowerMCP #CopilotStudio #MCP #PowerPlatform #AIAgents