Build a Custom MCP Server for Azure API Management in C#

🛠 Lab from the AI Platform Engineering on Azure short course. Doubles as the Azure-flavored extension for AI Platform Engineering Bootcamp Week 4 (the Python-leaning `ai-mcp-server` lab). Bootcamp landing page: https://academy.tekanaid.com/bootcamps/ai-platform-engineering-bootcamp Parent course(s):

• AI Platform Engineering on Azure (slug: ai-platform-engineering-azure)

🟡 Beta lab. Hands-on instructions, check scripts, and solve scripts are in active development.

Section 2 of the parent course made three design decisions: which slice of APIM's management surface is safe to expose as MCP tools, how to build an MCP server in C# with the official .NET SDK, and how to keep the tool surface read-only and token-efficient. This lab is where those decisions become a working artifact.

You provision a real Azure API Management instance on the Consumption tier (~5-minute provision, $3.50/M calls — no mock), seed it with two demo APIs, then scaffold a C# console project on .NET 8 that hosts an MCP server via `ModelContextProtocol` 1.3.0 + `ModelContextProtocol.Core` 1.2.0 + `ModelContextProtocol.AspNetCore` 1.2.0. The server starts on stdio so it can be driven by Claude Desktop, the MCP Inspector, or a local agent process. You ship a `ping` smoke-test tool first, verify the JSON-RPC handshake by hand, then implement the four lesson-defined read tools: `apim_list_apis`, `apim_get_api_config`, `apim_get_policies` (summarized — no raw XML), and `apim_get_traffic_metrics`. Authentication is service-principal via `DefaultAzureCredential` and the `Azure.ResourceManager.ApiManagement` SDK.

The final two tasks wire the server into real clients. Task 4 connects MCP Inspector and walks through a natural-language interaction where the LLM picks the right tool and summarizes the response. Task 5 spins up a slim Microsoft Agent Framework ReAct agent (mirroring the Lab 1 pattern, trimmed for this lab's VM) that wires both the official `Azure.Mcp` server from Lab 1 AND your new APIM MCP server side-by-side. You send a single prompt that touches both — "what AKS clusters do I have and what APIs are in our APIM instance?" — and watch the agent fan out to two MCP servers, fuse the responses, and answer in one shot.

By the end you will have built and operated the Azure-specific MCP-server pattern Lab 3 needs: a read-only, token-efficient, service-principal-authenticated MCP surface over a critical platform-engineering control plane, consumed from both a desktop client and an agent in the same loop.