MCP

📘What it really is

• The Model Context Protocol (MCP) is an open standard (introduced by Anthropic in 2024 and now widely supported) for connecting AI applications to tools and data sources.

• Parts:

• an MCP server offers tools (actions), resources (data to read) and prompts (templates); for example, an MCP server for your CRM, GitHub, a database or a file system;
• an MCP client inside an AI application (chat apps, coding assistants, your own app) connects to servers and makes their tools available to the model.
• Benefits: write a tool integration once, and use it from many AI applications; many ready-made servers exist.

• Transports: a local server runs on your machine (communicating through standard input and output); a remote server runs over HTTP, usually with OAuth for authentication.

• Security is your responsibility:

• only install MCP servers from sources you trust, and read what tools they expose; a server runs code and sees data;
• the same rules as Lesson 3: least privilege, read-only by default, the user's own permissions, approvals for actions;
• watch for tool descriptions or results containing hidden instructions (a form of prompt injection);

• give remote servers scoped, revocable tokens, never your main credentials.

• For a single app with a few internal tools, direct tool calling (Lesson 1) is simpler. MCP shines when tools should be shared across several AI applications.