What MCP Actually Is — a Shared Plug Shape
Last part ended on a gap: an agent in your terminal can only reach things that have a command on your machine. Your ticket tracker, a hosted database, your company wiki — those live behind a server and an API, the two words from Part 9. Getting an AI to them was, until recently, a genuinely bad problem.
The problem was multiplication
Say there are five AI apps and twenty systems worth connecting. That is not twenty-five pieces of work. Each app needs its own custom glue for each system, so it is five times twenty — a hundred. And every one of those hundred separately reinvents signing in, handling failures, and describing what the tool can do. Badly, usually, because nobody has time to do a hundred of anything well.
This is why so many obvious integrations never existed. Not because they were hard, but because a hundred pieces of near-identical work is nobody’s job. MCP replaced the multiplication with addition: agree on one shape, and each side builds to it once.
Two sides, plainly
- The client is the AI app you are sitting in front of. Your coding agent, a desktop assistant. It knows how to speak MCP and nothing about your particular systems.
- The server is a small program that exposes exactly one system — your files, a database, a ticket tracker. It knows that system properly and nothing about which AI is asking.
Server here has nothing to do with a data centre. It is usually a small program running on your own laptop. Same word as Part 9, same plain meaning: something that waits to be asked and answers.
What a server offers: three things
- Tools are named actions the model can take — create a ticket, run a query. Each one carries a description the model reads and a declared list of what it needs, so the model can discover what is possible instead of being told in advance.
- Resources are things to read, sitting at an address — a file, a record, a page. They inform an answer; they do not change anything.
- Prompts are good instructions saved next to the tool, so the sensible way to use it travels with it instead of living in one person’s head.
Notice what the tool description is doing. It is the same idea as --help in Part 5, pointed at a model instead of a person: the tool explains itself, so nobody has to hard-code a list of what exists. That is why a client written last year can use a server written tomorrow.
Adding one, in practice
claude mcp add --transport http sentry https://mcp.sentry.dev/mcpclaude mcp add my-server -e API_KEY=xxx -- npx my-mcp-serverclaude mcp listThat last command is worth running before you debug anything, because it tells you the difference between a server you configured and a server that is answering. Mine lists several, and some of them say they need authentication — which is the subject of the final part.
Which brings us to the finale, and to the most over-built thing in AI right now: an MCP server for something a plain command already does perfectly well.