Under the hood | Master AI Automation in 4 hours Master AI Automation in 4 hours Course About Ayush Modules Sample chapter Toolbox The Microcap Minute Classroom / Module 05: MCP: Plug-ins With Standards / Chapter 2 Under the hood Watch first, then read. Same lesson, your pace. What you will learn – The three things any MCP server offers – How a conversation flows through the protocol – How to judge a server before installing it A menu of three courses Every MCP server offers some mix of three primitives: Primitive What it is Example from a filesystem server Tools Actions the AI may perform read_file , write_file , list_directory Resources Data the AI may read File contents exposed as readable documents Prompts Ready-made templates the server ships “Explain this codebase” starter Tools are the stars, verbs the model chooses when a task needs them. When you say “clean up my downloads”, a filesystem tool gets called under the hood. Resources feed context; prompts package know-how. A round trip, step by step Where does intelligence live? Nowhere in the server, and that’s the elegant part: You tell the client (e.g., Claude): “What’s in my projects folder?” Client asks its connected servers: “what tools do you have?” → server answers: list_directory , read_file … (a plain description, no execution) The model decides which tool fits and emits a request: call list_directory(“/Users/ayush/projects”) Client executes via the server → results return Model reads results, answers you, possibly chaining more calls So the server is hands, the model is judgement, the client is the referee. This loop, decide, act, observe, repeat, is quietly the definition of an agent , and MCP is how today’s agents get hands. Reading a server before trusting it Servers are usually npm/pip packages or GitHub repos, and quality varies wildly. Before installing anything, check three things: Who publishes it? Official orgs (Anthropic’s reference servers, GitHub’s own) > random accounts. What exactly do its tools claim? A server wanting write_file everywhere when you only need reads is over-permissioned. Is it maintained? Last commit recent? Issues answered? You’ll apply this checklist for real in Chapter 3, reading first is what separates confident users from victims. Try it yourself Visit github.com/modelcontextprotocol/servers (the official reference collection). Pick any server and find in its README: its tool names, whether it exposes resources, who maintains it. Then explain the round trip above aloud to yourself using that server as the example, “if I asked X, the model would call Y”. If you can narrate it, you understand MCP better than most professionals. Key takeaways – Servers offer tools (actions), resources (data), prompts (templates). – The model decides; the server acts; the client referees, intelligence and hands stay separate. – Tool-calling loops are agents in embryo. – Vet servers like apps: publisher, permissions, maintenance. Download the exercise sheet (PDF) Module workbook (PDF) ← Prev: The USB-C moment Next: Plug in your first server → Classroom / Module 05: MCP: Plug-ins With Standards / Chapter 2 Under the hood What you will learn – The three things any MCP server offers – How a conversation flows through the protocol – How to judge a server before installing it A menu of three courses Every MCP server offers some mix of three primitives: Primitive What it is Example from a filesystem server Tools Actions the AI may perform read_file , write_file , list_directory Resources Data the AI may read File contents exposed as readable documents Prompts Ready-made templates the server ships “Explain this codebase” starter Tools are the stars, verbs the model chooses when a task needs them. When you say “clean up my downloads”, a filesystem tool gets called under the hood. Resources feed context; prompts package know-how. A round trip, step by step Where does intelligence live? Nowhere in the server, and that’s the elegant part: You tell the client
Under the hood
Written by
in