MCP Server

MCP Server

Overview

MCP is an open protocol that standardizes how applications provide context to LLMs and is often called the "USB-C port for AI Applications". MCP provides a standardized way to connect AI models to different data sources, prompts and tools.

An MCP (Model Context Protocol) server provides Resources, Tools, and Prompts to an application, enabling it to enhance its communication with a Large Language Model (LLM) or Small Language Model (SLM). The application initiates the MCP server—typically within its own security context—and communicates with it via standard I/O (stdio)

An application is not limited to a single MCP server; it can start and interact with multiple MCP servers simultaneously, each providing distinct Resources, Tools, and Prompts. While there is a one-to-one relationship between each MCP client (initiated by the application) and its corresponding MCP server, the application can manage several such pairs. In the given example, three MCP servers are launched, each contributing their own contextual assets to enhance model interactions.

Tools: Often used to provide functionality to the application (e.g. retrieve private information, perform specific tasks etc.). A list of available MCP server tools can be provided by the app to the LLM/SLM instance (with function calling capabilities) where most SDKs wrap the necessary function or tool calls to the MCP server automatically.

Resources: Data which the app can request from the MCP server to e.g. ground LLM/SLM calls. The app can also subscribe to a MCP resource to get continous updates (if new information is available). The MCP server implements the necessary update logic.

Prompts: The MCP server can provide prompt templates for the app to send to LLMs/SLMs.

Repo Content

The repo contains a simplified c# MCP Server and a c# console application consuming Tools, Resources and and a script to setup an Azure OpenAI instance with a deployed gpt-4 model instance:

Folder	Appliation
./src/SportApp	A simplified c# console application which uses Tools, Prompts and Resources from SportKnowHow_McpServer
./src/SportKnowHow_McPServer	A simplified c# Mcp Server
./setup/setup.azcli	A powershell Azure CLI script to setup the necessary Azure OpenAI instance with deployed gpt-4 instance

Demo Application

Overview "SportsApp"

• SportApp starts the provided MCP server using dotnet run and stdio

      //MCP Server: start parameter
      string command = "dotnet";
      string[] arguments = new string[] {
          "run",
          "--project",
          "./src/SportKnowHow_MCPServer/SportKnowHow_MCPServer.csproj"};
  
      //MCP Server: communication protocol
      StdioClientTransportOptions stdioClientTransportOptions = new StdioClientTransportOptions()
      {
          Name = "Sport Know How (MCP Server)",
          Command = command,
          Arguments = arguments,
      };
      StdioClientTransport stdioClientTransport = new StdioClientTransport(stdioClientTransportOptions);
  
      //Start MCP Server
      await using IMcpClient mcpClient = await McpClientFactory.CreateAsync(stdioClientTransport);
  

• All available tools provided by the MCP server are listed

• The tool with id RetrieveWinner with parameter eventName' and eventDate` is executed

      //List available tools
      IList<McpClientTool> mcpClientTools = await mcpClient.ListToolsAsync();
   
      //Manual tool call
      string toolName = "RetrieveWinner";
      Dictionary<string, object?> callParameter = new Dictionary<string, object?>
      {
          { "eventName", "Super Sports Champtionship" },
          { "eventDate", "2025-01-01"}
      };
      CallToolResponse callToolResponse = await mcpClient.CallToolAsync(toolName, callParameter);
  

• The MCP server tools are provided to a LLM/SLM completion call using Microsoft.Extensions.AI. There's no need to intercept the LLM/SLM function tooling requests as chatClient.RetResponse() takes care of calling the MCP Server tools as requested by the LLM and provides results back to the LLM.

      //Use tools in chat completions
      ...
      string systemMessage = "You help with information around international sport events.";
      string userMessage = "Who won the Super Sports Championship 2025?";
      var chatMessages = new List<ChatMessage>
      {
          new ChatMessage (Microsoft.Extensions.AI.ChatRole.System, systemMessage),
          new ChatMessage (Microsoft.Extensions.AI.ChatRole.User, userMessage)
      };
      ChatResponse chatResponse = await chatClient.GetResponseAsync(
          chatMessages,
          new ChatOptions
          {
              Tools = mcpClientTools.ToArray()
          }
      );
      ...
  

• Within the MCP server defined prompts are listed and a specific prompt (CreateMatchSummaryPrompt) with parameters is requested.

      // List available prompts
      IList<McpClientPrompt> mcpClientPrompts = await mcpClient.ListPromptsAsync();
      
      // Retrieve specific prompt
      string promptName = "CreateMatchSummaryPrompt";
      Console.WriteLine($"Retrieve prompt: {promptName}");
      Dictionary<string, object?> promptParameter = new Dictionary<string, object?>
      {
          { "eventName", "Munich Flying Dolphins vs. Berlin Bears" },
          { "winner", "Munich Flying Dolphins" },
          { "score", "24:31" },
          { "enthusiasm", "high" }
      };
      GetPromptResult getPromptResult = await mcpClient.GetPromptAsync(promptName, promptParameter);
  

• Within the MCP server defined resources templates and resources are listed and requested.

      //List available resource templates
      string resourceTemplateUri = "";
      IList<McpClientResourceTemplate> mcpClientResourceTemmplates = await mcpClient.ListResourceTemplatesAsync();
      
      //Retrieve a specific resource template
      resourceTemplateUri = resourceTemplateUri.Replace("{id}", "NurembergFlyingTigers"); 
      ReadResourceResult readResourceTemplateResult = await mcpClient.ReadResourceAsync(resourceTemplateUri);
      
      //List available resources
      string resourceUri = "";
      IList<McpClientResource> mcpClientResources = await mcpClient.ListResourcesAsync();
      
      // Retrieve a specific resource
      ReadResourceResult readResourceResult = await mcpClient.ReadResourceAsync(resourceUri);
  

• A subsription with the MCP server is created. The MCP server can send updates whenever the underlying resource is changed.

      // Subscribe to resources
      mcpClient.RegisterNotificationHandler("notifications/resource/updated", NotificationHandler);
      await mcpClient.SubscribeToResourceAsync("sportIntelligence://team/MunichFlyingDolphins");
  

Overview "SportKnowHow_McpServer

• MCP Server creation with Tools, Prompts and Resource support.

  //create MCP server
  HostApplicationBuilder hostApplicationBuilder = Host.CreateApplicationBuilder(args);
  hostApplicationBuilder.Logging.AddConsole(consoleLogOptions =>
  {
      consoleLogOptions.LogToStandardErrorThreshold = LogLevel.Trace;
  });
  
  hostApplicationBuilder.Services
      .AddMcpServer()
      .WithStdioServerTransport()
      .WithTools<SportEventTools>()
      .WithPrompts<SportEventPrompts>()
      .WithResources<SportEventResources>()
      .WithSubscribeToResourcesHandler(SubscribeToResourcesHandler)
      .WithUnsubscribeFromResourcesHandler(UnsubscribeFromResourcesHandler)
      .WithReadResourceHandler(ReadResourceHandler);
  
  hostApplicationBuilder.Services.AddSingleton(resourceSubscriptions);
  hostApplicationBuilder.Services.AddHostedService<SubscriptionMessageSender>();
  

Summary

MCP Server which often exectue in the security context of the calling applications make it easy to provide tools,resources and prompts to calling applications. Just like shown in the sample SportApp application the ModelContextProtocol nuget package simplifies the development of Mcp Server and Mcp Host applicatons.