Skip to main content
Glama

Introduction to Windows AI Foundry & MCP on Windows 11

Written by on .

Windows
mcp
Foundry
  1. What It Means for Developers
    1. Implementation (C# minimal MCP server)
      1. Behind the Scenes
        1. My Thoughts
          1. References

            In May 2025, Microsoft introduced Windows AI Foundry alongside native support for the Model Context Protocol (MCP) on Windows 11. This update allows AI agents to connect with Windows components such as the file system and the Subsystem for Linux (WSL) in a standardized way 1. The feature was announced at Build 2025 and marks a significant step in enabling AI applications to function locally within Windows without complex integrations 23.

            What It Means for Developers

            Windows AI Foundry replaces the earlier Copilot Runtime and provides developer tools for fine-tuning models and running them directly on Windows 11 devices. By integrating MCP at the operating system level, Windows exposes key system functions through a registry of MCP servers. Developers can now build agents that discover these servers and invoke tools like file system access or window management 23. This means agents can perform tasks such as “search my documents folder” within Windows, with fewer steps and greater efficiency.

            Image

            Implementation (C# minimal MCP server)

            Here is a basic C# example of creating a minimal MCP server on Windows:

            using Microsoft.Mcp.Server; using Microsoft.Mcp.Transport.Stdio; var host = new McpServerHostBuilder() .WithStdioServerTransport() .Build(); host.RegisterToolAssembly(typeof(MyFileSystemTool).Assembly); host.Start(); public class MyFileSystemTool { [McpTool("list_files")] public string[] ListFiles(string directory) { return System.IO.Directory.GetFiles(directory); } }

            In this example, the MCP server runs over standard input/output, registers a tool assembly, and exposes a list_files method. A compatible agent can locate it through the Windows MCP registry and invoke the tool 45.

            To configure an Azure AI Foundry agent to use this server:

            "tools": [ { "tool_type": "mcp", "server_url": "https://localhost:5000", "server_label": "local_fs", "allowed_tools": ["list_files"], "require_approval": "always" } ]

            This agent configuration ensures the agent discovers the server, only uses specified tools, and prompts for user consent before each call 6.

            Behind the Scenes

            When you launch an MCP server on Windows 11, it registers with the system MCP registry, which maintains a verified list of available servers. When an agent queries for tools, it performs discovery through this registry. The user is prompted for approval in a UAC-style dialog before each sensitive tool call. Once approved, agents use JSON-RPC over HTTP or Stdio transport to call tool functions. Windows enforces registry trust, tool isolation, and logs permissions to monitor usage and prevent misuse 32.

            Image

            My Thoughts

            This integration simplifies agent development on Windows and reduces friction around agent-tool discovery. The registry model and user approval flow balance flexibility with security. However, developers should monitor the frequency of approval prompts to avoid user fatigue. Enterprises may need to manage directories of approved MCP servers. Overall, this is a strong step toward making Windows a more agent-friendly platform for developers.

            References

            Footnotes

            1. The LetterTwo article on Windows AI Foundry and MCP support ↩

            2. The Verge article on MCP support designed for security and registry control ↩ ↩2 ↩3

            3. Directions on Microsoft Build coverage citing MCP across services ↩ ↩2 ↩3

            4. Microsoft Learn guide to Azure AI Foundry MCP client integration ↩

            5. Wikipedia page on the Model Context Protocol architecture and adoption ↩

            6. Microsoft Learn code samples and agent configuration for MCP tools ↩

            Written by Om-Shree-0709 (@Om-Shree-0709)