MCP

MCP (Model Context Protocol) is a game-changing technology that is all the hype right now, but I have heard some confusion around it. Here’s how I think about it.

An example of MCP is letting Claude desktop read a file on your computer with instructions, write some code, commit that code to Github, and open a PR. MCP is a way for an LLM to discover and use your tool or service.

Let’s say you have a website that looks for flights on travel sites. You want customers to be able to type “help me find a flight for this weekend from LAX to Austin nonstop” into ChatGPT (or any AI chat) and have it search your website and return the results. If they find a flight they like, book it all through the chat.

MCP allows that to happen. It’s a way to expose your functionality to LLMs, in our example it’s searching for flights and booking them. Creating an MCP server allows an LLM to discover what it can do with your service, and then execute those tasks.

If you are creating an AI tool that has tools you don’t necessarily need to create an MCP server. You can just execute tools with a normal API request. An MCP server is useful for exposing your tools to multiple agents, or for code separation and standardization with large teams.