Practical MCP with FastMCP & LangChain

Why MCP Matters?

Consider a common scenario: you want your AI assistant to create an issue on GitHub, upload a file to Dropbox, and send a message in Discord. Without MCP, the LLM needs to understand three completely different APIs. Each service has its own authentication method, its own request format, its own error handling, and its own quirks. The application developer has to write and maintain three separate integrations, and if a fourth service comes along, that means a fourth custom adapter.

The diagram shows this contrast clearly. On the left side, the LLM connects directly to three different APIs: GitHub, Dropbox, and Discord. Each connection is its own integration with its own logic. The application has to know the details of every service it talks to.

On the right side, the same three services are available, but the architecture is fundamentally different. The LLM speaks a single unified API, the Model Context Protocol, through an MCP client. That client connects to an MCP server, and the server is the one that knows how to talk to GitHub, Dropbox, and Discord. The LLM does not need to understand the specifics of any of those services. It only needs to speak MCP.