Course 2 - Module 3 - Lesson 1
Tools, resources, prompts, clients, servers, and permissions explained with a kitchen-menu model
MCP: The Tool Menu
Who talks to whom
Model Context Protocol, or MCP, standardizes communication between AI applications and connected capabilities. The host is the AI application. A client inside that host communicates with an MCP server. The server exposes particular capabilities; it is not necessarily the place where your language model runs.
Picture a restaurant. The host application coordinates your visit, a client carries requests, and a server presents a menu of available capabilities. This analogy helps with roles, but the real system still needs explicit connection settings, access controls, and careful handling of returned data.
Three things on the menu
Tools are callable functions, such as looking up a booking or creating a draft. Resources provide contextual information, such as a document or database description. Prompts are reusable instruction templates. A server may offer some of these features without offering all of them.
In the restaurant picture, a tool places a defined request, a resource provides the menu's contents, and a prompt gives a reusable way to ask for meal suggestions. A lookup tool may only read information; tool does not automatically mean a write or an external change.
A connection is not blanket permission
Suppose a practice server exposes a fictional cafe menu and a draft-order tool. Reading the menu should not imply permission to place a paid order. The application and connected service must enforce which operations are allowed and where human approval is required.
MCP supplies a communication structure, not a promise that every server is trustworthy or every integration compatible. Review the server's identity and requested access. Treat text returned by a tool as information to evaluate, especially if it tries to issue unrelated instructions.
Tiny experiment
For a fictional cafe assistant, name one tool, one resource, and one prompt. Mark which proposed action would require approval before money is spent.
Confidence check
Why would a shared communication protocol still need separate rules about identity, access, and user consent?