Model Context Protocol #

Generally, I try to learn how something works before using it. However, in the case of MCP, I used it and enjoyed the benefits of it before I thought that I should start learning about it.

Specifically, I have used the GitHub MCP and the Atlassian Confluence one, which has been great for reviewing code and updating docs.

This page will document my learnings about MCP.

Why? #

MCP standardizes the way that AI systems connect models with the tools that are required to perform tasks. Something like how LSP servers made it easier for multiple IDEs to get deeper insights into multiple programming languages.

If it’s an API based tool e.g., GitHub PR API, why wouldn’t you let the AI agent call the API directly? That wouldn’t be efficient, given the lack of context. Maybe the current API themselves weren’t made with AI agents in mind and return responses in an inefficient manner. An MCP server could handle responses better.

What is it made of? #

I’m going to talk about an MCP server first. It’s made of :

1. Tools. Which can be used by clients to perform actions on the server.
2. Resources. Which the AI agent can read to provide answers to the user.
3. Prompt template. Tells how to write prompts to get the best response from the server.

How does it work? #

There is a lifecyle in the communication between an MCP server and client.

1. Initialization
2. Message Exchange (through Streamable HTTP. Which will use normal HTTP REST APIs but can upgrade to SSE for push)
3. Termination