AI & LLMs
MCP servers, explained simply
Imagine the smartest kid in school. She knows a ton — but she's sitting in a room with no phone, no calculator, no window. Ask her today's weather and she can only guess. That's a large language model on its own: brilliant, but with no hands.
MCP — the Model Context Protocol — is how we give that kid a set of doors. Each door leads to one useful thing: a weather service, your files, a database, a calculator. Open standard, same door-handle everywhere. Let's look at the picture first, then the plumbing.
MCP is like a universal USB port for AI. Before USB, every gadget had its own weird plug. USB made one shape that fits everything. MCP is that one shape — but for connecting an AI to tools.
So what actually is it?
MCP is an open protocol — a shared rulebook — introduced by Anthropic for connecting AI assistants to the systems where data and tools live. Instead of every app inventing its own custom glue for every tool, everyone agrees on one format. Build a tool once as an "MCP server," and any MCP-aware app can use it.
There are three roles worth knowing:
- Host — the app you're actually using (a desktop assistant, an IDE, a chat client). It wants to get things done.
- Client — a small connector living inside the host. It speaks MCP and manages one connection per server.
- Server — the program that exposes a capability: a weather server, a filesystem server, a GitHub server. It says "here are the tools I offer" and does the work when asked.
The flow, step by step
Walk it through with a question like "what's the weather in Pune?":
- The host gives the model your question along with a list of tools its MCP servers offer.
- The model decides it needs the
get_weathertool and asks for it. - The client forwards that request over MCP (a JSON-RPC message) to the weather server.
- The server calls the real weather API, gets "31°C, humid," and hands it back.
- The model reads the result and writes you a normal sentence.
Why this is a big deal
- Write once, use everywhere. A single filesystem server works in any MCP-aware app — no bespoke integration per app.
- Swappable. Because the "plug" is standard, you can add or remove tools without touching the model.
- Boundaried. The server decides exactly what it exposes, so the AI only gets the specific doors you choose to open — not the whole house.
A language model is a brain in a jar. MCP is the standard set of arms you bolt on — and every arm clicks into the same socket.
That's the whole idea. Next time you see an assistant check your calendar, search a codebase, or file a ticket, there's a good chance a small MCP server is doing the reaching — while the model just decides which door to open.