Course B · Module B2 Going Agentic · Lesson 11 of 18

Tools, CLI, MCP & Skills

Four ways to give your agent superpowers — and how to know which one a job needs.

~13 min · Have your terminal open · Prereq: B10 Anatomy of an Agent

In B10 you learned that an agent needs "tools" — but nobody explains what that means in practice, or why the words CLI, MCP, and skills keep showing up alongside it. By the end of this lesson you'll be able to decode all four — and, crucially, you already have hands-on experience with two of them: at the very start of this course you installed the teach skill by typing a CLI command (npx skills add). Your single win: understand what each layer does and when to reach for it, so you can confidently wire up an agent that reaches the outside world.

Tools: specific abilities you grant

A tool is a defined capability the model can call during its loop. You grant tools deliberately — the agent can only do what you've unlocked. Common examples:

web_searchLook something up on the internet in real time.

send_emailCompose and deliver an email via your email provider.

read_fileOpen and read a file from your computer or cloud storage.

query_databaseRun a read or write against your data store.

run_codeExecute a script and get the output back.

The model decides when to call a tool and what to pass it. The result comes back like any other text and the model reads it, then decides what to do next. You are always in control of which tools exist — that's your guardrail lever.

CLI: let the agent use the command line

A CLI (command-line interface) is the text-based way to control software by typing commands instead of clicking buttons. It looks intimidating to humans — but for an agent it's one of the biggest superpowers of all. Almost every serious tool on a computer has a command line, and the moment you give an agent a terminal, that entire universe of existing tools becomes usable with no special integration to build.

Think of the CLI as handing the agent the same keyboard a developer uses. Suddenly it can do what a developer does by typing: save to git, deploy a site, resize a folder of images, run tests, install a package.

You've already done this. npx skills add was a CLI command — you typed it, but an agent could run that exact line itself. This is literally how coding agents like Claude Code work: they're given a terminal and they drive real tools the way a person would.

MCP: the universal plug standard

Before USB, every device came with its own proprietary cable. Printers had one connector, cameras had another, keyboards had yet another. Then USB arrived: one standard plug, any device. MCP — Model Context Protocol — is USB for AI agents.

Before MCP, connecting an agent to a new data source (your calendar, your CRM, a database) required custom code each time. MCP defines a single open standard so that any agent that speaks MCP can plug into any tool or data source that also speaks MCP — with no custom wiring.

What MCP gives you

One integration, any agent — write the connector once
Growing ecosystem of ready-made MCP servers (GitHub, Slack, databases, APIs)
Standardized security model — you configure what's exposed
Works across Claude, other models, and frameworks

What MCP is not

Not magic — you still decide which MCP servers to connect
Not a no-code platform — still some config required
Not automatic permission — you grant access intentionally
Not required for simple agents that use built-in tools only

In plain English: MCP means you can pick from a growing library of pre-built connectors instead of building from scratch. Connect your agent to Google Drive, Notion, your CRM, or your database by pointing at an MCP server — no custom glue code.

Skills: packaged expertise the agent loads on demand

A skill is a bundle of instructions — a SKILL.md file plus any supporting prompts — that tells an agent how to do something well. You've already used one. When you ran npx skills add at the very start of this course and installed the teach skill, you added a folder that contains a SKILL.md with instructions on how the agent should behave as a teacher: the tone to use, the structure of explanations, when to ask you questions back. The agent loads that file into its context and suddenly "knows" how to teach.

SKILL.mdThe instructions file — plain English directions telling the agent how to approach a task. This is the whole skill.

npx skills addThe command to install a skill from the registry into your project.

/teach (or any /command)How you invoke a skill once installed — a slash command that loads the relevant SKILL.md.

Skills are composable: one agent can have several. A "research-and-write" agent might load a research skill (thorough sourcing, skepticism) and a write skill (your brand voice, paragraph length). Each skill loads its instructions into context at the right moment.

How the four compare

SkillsExpertise — instructions that teach the agent HOW to think about a task. Reach for it when: you want consistent, high-quality judgment.

ToolsA single defined ability the agent can call. Reach for it when: the agent needs to DO one specific thing.

CLIThe agent runs existing terminal commands. Reach for it when: a capable command-line tool already exists and you want power fast.

MCPA universal standard that plugs tools and data into any agent. Reach for it when: you want a clean, reusable, permissioned connection to a system.

CLI vs MCP — the key contrast: both connect an agent to the outside world, but differently. A CLI is immediate and enormously powerful — the agent borrows tools that already exist — but it's free-form: little structure, and it runs real commands on your machine, so guardrails matter. MCP is more deliberate — a structured, permissioned connector someone has built — safer and reusable, but only useful if an MCP server for that system exists. Rule of thumb: reach for a CLI when the tool is already there and you want power now; reach for MCP when you want a clean, repeatable, controlled connection.

Check yourself

Answer before opening — recalling it now is what makes it stick.

You want your agent to look up contacts in your CRM. Is that a tool, MCP, or a skill?

A tool — specifically, one delivered via an MCP server. "Query CRM" is a capability (a tool). If your CRM has an MCP server, you connect it via MCP so you don't have to write the connector yourself. Tool is the what; MCP is the how it's wired up.

What did the 'teach' skill you installed actually add to your project?

A folder containing a SKILL.md file — plain-English instructions telling the agent how to behave as a teacher. When you invoke the skill, the agent loads those instructions into its context. No code, just structured guidance.

Why does MCP matter to a non-engineer entrepreneur?

Because it means pre-built connectors exist for most common tools. Instead of paying a developer to wire your agent to Google Drive or Notion, you find (or run) an MCP server for that service and point your agent at it. The standard eliminates the custom plumbing cost.

Your agent needs to resize a few hundred images, and a command-line tool for that already exists. Fastest path — CLI or MCP?

Give the agent CLI access and let it run that tool. When a capable command-line tool already exists, the agent can just use it — nothing to build. You'd reach for MCP instead when you want a standardized, permissioned connection you'll reuse across many agents and systems.