OpenAI Adopts the Model Context Protocol, Starting With the Agents SDK

Why It Matters for Developers#

The practical effect is "write once, expose everywhere." A tool published as an MCP server can now be consumed by OpenAI agents, by Claude and by the growing list of editors and assistants that speak the protocol. For teams that maintain internal tools, the question changes from "which assistant do we integrate with?" to "how do we operate one well-secured MCP server?"

.NET developers can take part on both sides of the protocol with the official C# SDK (see our MCP in C# guide and the C# SDK launch news). Because the SDK's client tools plug into Microsoft.Extensions.AI, an MCP server can supply tools to any IChatClient, including one backed by an OpenAI model. The following sketch follows the pattern of the SDK's own ChatWithTools sample:

using Microsoft.Extensions.AI;
using ModelContextProtocol.Client;
using OpenAI;

// Start a local MCP server over stdio (here, the MCP project's reference test server).
var mcpClient = await McpClient.CreateAsync(
    new StdioClientTransport(new()
    {
        Name = "Everything",
        Command = "npx",
        Arguments = ["-y", "@modelcontextprotocol/server-everything"],
    }));

var tools = await mcpClient.ListToolsAsync();

var apiKey = Environment.GetEnvironmentVariable("OPENAI_API_KEY")!;
using IChatClient chatClient = new OpenAIClient(apiKey)
    .GetChatClient("gpt-4o-mini") // any tool-capable model
    .AsIChatClient()
    .AsBuilder()
    .UseFunctionInvocation()
    .Build();

List<ChatMessage> messages = [new(ChatRole.User, "Add 17 and 25 using your tools.")];
var options = new ChatOptions { Tools = [.. tools] };

await foreach (var update in chatClient.GetStreamingResponseAsync(messages, options))
{
    Console.Write(update);
}

A few design choices deserve attention:

• Pick the transport deliberately. Use stdio for local, per-user tools and Streamable HTTP for shared remote servers. Avoid building anything new on the deprecated SSE transport.
• Treat hosted calls as internet-facing. A server that OpenAI's platform calls directly must be publicly reachable, so it needs real authentication, rate limiting and least-privilege access to back-end systems.
• Keep a human in the loop for side effects. Approval policies and tool filtering exist for a reason. Tool descriptions and results are untrusted input that can carry prompt injection, as our responsible AI guide explains.
• Keep your abstractions portable. Coding against IChatClient, as covered in the Microsoft.Extensions.AI guide, lets you swap model providers without rewriting tool plumbing. The function calling guide covers the underlying mechanics.