Loading video player...
Lecture 21 of the AI for Software Engineers series ā Bipin Kumar explains MCP, the universal standard for connecting tools to LLMs. This is the concept that lets Claude access your Gmail, GitHub manage your repos through AI, and any team's tools become reusable by everyone. š§ What's Covered: Why MCP Exists ā The Redundancy Problem: Three teams in the same company all need to integrate Gmail and WhatsApp. Without MCP, each team writes its own version of the same connection code independently. With MCP, one team writes it once, hosts it on a server, and all three teams simply connect to it. No duplicate work. MCP Architecture ā 3 Parts: Server ā where the tools actually live. This is a separate file (tool.py) that can run on your local machine or on any cloud server. Earlier, tools lived in the same file as your LLM code. MCP separates them. Host ā your main application (app.py). This is the bigger program that contains everything. Client ā a small piece of code inside the host that handles connecting to the server. The client sits inside the host and is responsible for the connection only. Two Transport Protocols: STDIO (Standard Input/Output) ā used when the server runs on your local machine. The client automatically starts the server process. One-to-one connection only ā if two scripts try to use the same local tool at the same time, only one can. HTTP-SSE (Server-Sent Events) ā used when the server runs on the cloud or needs to serve multiple clients simultaneously. The server must be started manually. Multiple clients can connect to the same server at once. SSE means the connection stays open and the server keeps pushing updates ā like a food delivery status that updates automatically without you refreshing. FastMCP Library: The original MCP implementation required complex code with many lines. FastMCP is a wrapper library that makes writing MCP tools simple. The only change from a regular LangChain tool is replacing the decorator with the MCP version and adding a transport declaration at the end. MCP Marketplace: An ecosystem of pre-built MCP servers ā similar to an app store. Companies like Cloudflare, Perplexity, GitHub, Gmail, Zerodha, Neo4j, and Elasticsearch have all published official MCP servers. Connect to them without writing any integration code yourself. Created by Anthropic, Now Universal: Anthropic built MCP for Claude. It became so useful that the entire industry adopted it. When you give Claude permission to access your Gmail, it connects through MCP internally ā this is exactly the same protocol you are now learning to build with. What you can connect via MCP: custom functions, databases, APIs, documents, prompts ā anything. āļø Next Lecture (Lecture 22): š MCP Hands-On ā writing STDIO and HTTP tools with full code walkthrough š¬ Questions about STDIO vs HTTP-SSE? Drop them in the comments ā Bipin replies! š Subscribe so you never miss a class. #MCP #ModelContextProtocol #LangChain #AgenticAI #FastMCP #STDIO #HTTPSSE #AIforEngineers #BipinKumar #Anthropic #Claude #LLMTools #GenerativeAI #AIInterview #MCPServer