MCP stands for Model Context Protocol. It is an open standard that describes how an AI assistant talks to an outside service: how it discovers what tools exist, how it calls one, and what shape the answer comes back in.
The problem it solves is old and boring. Assistants are good at reasoning and bad at knowing current facts. Anything that changes — a price, a positioning reading, a calendar — is either stale in the model or absent from it. MCP gives the model a door to something that is current.
The three pieces
A server exposes tools. Each tool has a name, a description and a schema saying what arguments it takes. A client is the app you are using. The connection is usually just an HTTPS URL you paste into that app once.
When you ask a question, the client decides whether a tool is relevant, calls it, and reads the result back before answering you. You do not see the call unless the app shows it to you.
What it is not
It is not a plugin store, and it is not a way for a server to run code on your machine. The server answers questions; it does not reach into the client. A read-only server like a market-data one cannot change anything at all, and well-behaved tools declare that so the client knows it can call them without asking you every time.
It is also not magic about accuracy. A tool can only return what the server actually measured. The useful servers are the ones that distinguish between a value they measured, a value they retrieved from somebody else, and a value they do not have — and say which.
Why market structure suits it
Dealer positioning is computed, not reported. There is no headline that carries it, and it changes through the session. That makes it exactly the kind of thing a model cannot know and cannot look up in prose — and exactly the kind of thing a tool call can answer precisely.