MCP Basic Usage

1. Define tools in an XNCF module

Use attributes from ModelContextProtocol.Server on the tool type and methods:

using ModelContextProtocol.Server;
using System.ComponentModel;

[McpServerToolType]
public static class SampleMcpTools
{
    [McpServerTool, Description("Return the input unchanged")]
    public static string Echo(string message)
    {
        return message;
    }

    [McpServerTool, Description("Return the current server time")]
    public static string Now()
    {
        return DateTimeOffset.Now.ToString("O");
    }
}

Enable MCP in that module's Register:

public override bool EnableMcpServer => true;

The default scanner uses the assembly containing the module Register. If tools live in another assembly, override the registration explicitly rather than assuming the default scan crosses assembly boundaries.

2. Determine the module endpoint

The default route uses the complete module name:

Module:   Senparc.Xncf.MCP
Route:    mcp-senparc-xncf-mcp
Endpoint: http://localhost:5000/mcp-senparc-xncf-mcp/sse

At runtime, enumerate registered services with:

var servers = XncfRegisterManager.McpServerInfoCollection.Values;
foreach (var server in servers)
{
    Console.WriteLine($"{server.XncfName}: {server.McpRoute}/sse");
}

3. Use MCP from NCF AgentKernel

The current MCP module uses HostedMcpServerTool to expose a remote MCP server as an AI tool:

using Microsoft.Extensions.AI;
using Senparc.AI.AgentKernel;

var endpoint = new Uri(
    "http://localhost:5000/mcp-senparc-xncf-mcp/sse");

var mcpTool = new HostedMcpServerTool("NCF-Server", endpoint);
var handler = new AgentAiHandler(Senparc.AI.Config.SenparcAiSetting);
var config = handler.IWantTo();
var options = config.CreateChatClientAgentOptions(
    "NCF-Agent",
    "Call MCP tools when needed.",
    new ChatOptions
    {
        Instructions = "Call MCP tools when needed.",
        Tools = new List<AITool> { mcpTool },
    });

var runner = await config
    .ConfigChatModel("NCF-Agent", options)
    .BuildKernelAsync();
var result = await runner.RunChatAsync("Return the server time");
Console.WriteLine(result.OutputString);

A working chat model must already be configured. Do not disable human approval by default for tools that write data, execute commands, or access external systems.

4. Minimum verification

• Server logs show that the expected module MCP route was mapped.
• The client can list expected tools such as Echo and Now.
• Test one side-effect-free tool first, then add authentication, timeouts, and error handling.
• Production testing covers unauthorized access, timeout, cancellation, duplicate calls, and redaction of sensitive data.