A standard way to connect tools
The Model Context Protocol is an open standard for connecting Claude to external tools and data — "USB-C for AI." Instead of hardcoding each integration, a client speaks MCP to any server, and servers are reusable across apps.
Key concept · Clients, servers, and the three things a server exposes
An MCP client (inside your app or agent) connects to one or more MCP servers. A server exposes three kinds of capability: tools (actions Claude can call), resources (data/content Claude can read), and prompts (reusable prompt templates). You build a server by defining those, test it with the server inspector, then implement a client that discovers and calls them. Because it's a standard, a server you write once (or one someone else published) works with any MCP-speaking client — the whole point is dynamic, reusable tool discovery instead of bespoke wiring.
Claude Code and computer use as building blocks
Beyond raw API calls, Anthropic ships apps you can build with and around.
Key concept · Claude Code and computer use
Claude Code is an agent you can drive from the terminal — set it up, watch it work in a codebase, and enhance it with MCP servers so it reaches your tools and data (you can also parallelize it and use it for automated debugging). Computer use is a capability where Claude operates a computer through screenshots and mouse/keyboard actions — clicking, typing, navigating a UI — for tasks that have no API. Understanding how computer use works (a perceive-decide-act loop over screen state) tells you where it fits: the fallback when there's no cleaner integration, used carefully on trusted surfaces.
Patterns — and knowing which you need
The course closes by distinguishing two ways to build multi-step systems, and the patterns within each.
Key concept · Workflows vs. agents
A workflow follows a predefined path your code lays out; an agent decides its own path dynamically in the loop. Workflows are more predictable and debuggable; agents are more flexible but less predictable — start with the simplest thing that works, and only reach for an agent when the task genuinely needs dynamic control. Common workflow patterns: chaining (output of one step feeds the next), routing (classify the input, send it to the right handler), and parallelization (run independent subtasks at once and aggregate). Agents and tools plus environment inspection (letting the agent look at its surroundings before acting) round out how an agent operates.
Chapter 22 quiz
Fourteen questions on MCP, apps, and agents vs. workflows. Answer first, then expand Show answer.
Part A · Multiple choice
MCP is best described as:
- A An open standard for connecting Claude to external tools and data
- B A new model
- C A billing system
- D A prompt template language only
Show answer
A. "USB-C for AI."
What three things does an MCP server expose?
- A Tools, resources, and prompts
- B Models, keys, and logs
- C Chat, Cowork, Code
- D HTML, CSS, JS
Show answer
A.
What tests an MCP server as you build it?
- A The server inspector
- B A linter
- C The Console only
- D git diff
Show answer
A.
The advantage of MCP being a standard is:
- A A server written once works with any MCP-speaking client
- B It's faster to type
- C It removes the need for tools
- D Nothing
Show answer
A. Reusable, dynamic tool discovery instead of bespoke wiring.
Computer use is best understood as:
- A Claude operating a computer via screenshots and mouse/keyboard, for tasks with no API
- B A faster model
- C A database connector
- D A prompt technique
Show answer
A. A perceive-decide-act loop over screen state.
How can you extend Claude Code's reach?
- A Enhance it with MCP servers
- B Raise the temperature
- C Remove its tools
- D Use a bigger key
Show answer
A.
The difference between a workflow and an agent is:
- A A workflow follows a predefined path; an agent decides its own path dynamically
- B Workflows use no code
- C Agents are always cheaper
- D There is no difference
Show answer
A.
Part B · True or false
You should start with the simplest thing that works and only use an agent when needed.
Show answer
True. Workflows are more predictable and debuggable.
Routing means classifying the input and sending it to the right handler.
Show answer
True.
Chaining feeds the output of one step into the next.
Show answer
True.
Parallelization runs independent subtasks at once and aggregates the results.
Show answer
True.
MCP resources are actions Claude executes.
Show answer
False. Resources are data/content Claude reads; tools are the actions.
Part C · Short answer
Name the three workflow patterns.
Show answer
Chaining, routing, and parallelization.
When should you reach for an agent over a workflow?
Show answer
When the task genuinely needs dynamic, self-directed control — otherwise prefer the simpler, more predictable workflow.