MCP - Core Concepts

April 9, 2025

---

The world moves fast. The AI world moves 10x faster. Here are some notes I've taken recently while trying to wrap my head around the concept of MCP that appeared recently.

MCP who?

MCP stands for Model Context Protocol, a standard format used by the LLMs to interact. I like to think of it as API that is used exclusively by the LLM-based applications.

MCP why?

• The main advantage of the MCP is that the cognitive load needed to use the external data is no longer on the developers' shoulders but on the LLMs'.
• Users interacting with an LLM app can add new functionalities themselves. Those functionalities do not need to preexist.

MCP how?

• MCP server allows using the data from one app without the need of coding the logic upfront. It means that the devs working on LLM-based apps (clients) do not need to know how exactly the third-party apps provide their data (API) but rather they can let the LLMs figure it out themselves.
• The new features can be added in runtime, not in the compile-time.
• MCP server provides the context to the LLMs.
• The MCP servers can run either locally (stdio) or remotely (HTTP through SSE).
• They have only one entry point - their URL that allows to establish an SSE connection.
• MCP server can provide a list of tools that are ready to be used by the clients.
• A Tool in the context of an MCP executes some specific function. It can, for example, export the last month's data in a CSV format. The tools can be used by the LLMs.