AI Engineer’s Handbook to MCP Architecture

Author(s): Vatsal Saglani

Originally published on Towards AI.

Part 3: The Minimal MCP ClientBlog Header Generated by the Author

In the previous blog, we built two specialized servers: a File Management Server for document operations and a Run Command Server for shell execution. In this blog, we’ll develop the crucial bridge component, the MCP Client hub that connects language models to these servers.

We’ll be building this hub using FastAPI. Below, we provide a quick survey of existing MCP clients, starting with Claude Desktop’s native tool support. Then we’ll implement a minimal hub with a clean separation between discovery and execution. Our MCP Client hub will be built from first principles with an emphasis on simplicity and transparency, using FastAPI, Pydantic, and httpx.

Explore the entire codebase of the MCP Client Hub in the GitHub repo below:

Contribute to vatsalsaglani/nano-MCP development by creating an account on GitHub.

github.com

Anthropic’s Claude Desktop includes native MCP client capabilities. If we search the public MCP repos and Anthropic’s docs we will not find any reference code related to the MCP client being used in Claude Desktop.

Users enable “MCP Clients” in Labs settings, configuring servers via claude_desktop_config.json. Claude Desktop then handles tool discovery and routing over HTTP.

Several open-source alternatives have emerged:

mcp-use: Python package for MCP server interaction (pip install mcp-use)Dolphin-MCP:… Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI