Chapter 8: MCP — Give Your Agent Hands¶

Until now, your agent's world has ended at the edge of the project folder. MCP — the Model Context Protocol — is how that changes: an open standard that connects agents to everything else. Databases. Browsers. Design tools. Calendars. Your company's internal systems. Think USB-C for AI: one port, any device.

This chapter demystifies the protocol, gets your first servers connected, teaches the CLI alternative that pros often prefer, and — most importantly — builds your security judgment, because giving an agent hands means deciding what those hands may touch.

What you will learn

Explain why tools transform what agents can do, and what MCP standardizes.
Understand MCP's moving parts in plain language: servers, tools, resources.
Connect and use real MCP servers in Claude Code.
Use CLIs as agent tools — the underrated alternative that's often better.
Draw trust boundaries: which connections are safe, which need guardrails, which to refuse.

Builder principle

Every connection you give an agent is capability and attack surface. The pros aren't the ones with the most connections — they're the ones who can defend the ones they chose.

Watch and read with this chapter

MCP, Clearly Explained — the quick conceptual version.
Claude's Model Context Protocol — let's test it — Fireship's hands-on take, fast and irreverent.
Anthropic Academy: Introduction to MCP — free official course, the thorough companion.
modelcontextprotocol.io — the standard itself; bookmark.

Real talk¶

Connecting an agent to an actual system changes both its evidence and its risk. This chapter practices that decision with explicit capability, data, credential, and approval boundaries.

Lessons in this chapter¶

Chapter checkpoint¶

Complete Chapter 8: Least-privilege connection decision, save the evidence, then score your first attempt with the shared rubric.