MCP Servers & CLI

When to Use Which

GenAI Skills Academy

Friday 20 March 2026

Recap from Last Week

MCP = Model Context Protocol. An open standard that lets AI apps connect to external services.
Last week: built an SAP server live in Claude Desktop
Today: go under the bonnet. Understand the architecture, build another server, and compare MCP with CLI.

Three MCP Terms You Need to Know

Host (the app you type into)

Claude Code, Claude Desktop, VS Code, Cursor. Whatever AI app you're using. You already have this.

Server (the program you build)

A few files of code that know how to talk to an external API. Sits on your machine as a folder of files. Does nothing until the host starts it.

Client (the connector)

You don't build this. You never see it. Created automatically by the host when it reads your config. One per server.

What Do We Mean by "Server"?

In everyday language, a "server" is a machine in a data centre. In MCP, it means something different.

An MCP server is a program that translates between your app and an external service. It's software on your computer, not hardware in a rack.

Local server = a program you build, running on your machine. The connector (client) talks to it via stdio (standard input/output — how local programs talk to each other).
Remote server = a program Anthropic hosts for you. The connector talks to it over HTTP. This powers Google Drive in Claude Desktop.

Building an MCP Server

The Sequence

1

You already have the host (app) — Claude Code

2

We build the server (program) — code that calls the Google Drive API

3

Add it to your config file — tells the host where the program is

4

Restart Claude Code — creates a client (connector), connects to server, discovers tools

5

Ask a question — model picks the tool, connector calls server, server calls API

Diagram 1: Claude Code — Local Server

YOU ▶ CLAUDE CODE ▶ Claude model decides to call a tool (host — the app) | | tool call to connector v MCP CLIENT ▶ MCP SERVER ▶ GOOGLE DRIVE (connector) (program — (external your machine) service) JSON-RPC HTTP request over stdio to API

MCP (client ↔ server) happens on your machine. Anthropic's servers are only involved for the model's thinking — not part of MCP.

Diagram 2: Claude Desktop — Remote Server

YOU ▶ CLAUDE DESKTOP ▶ Claude model decides to call a capability (host — the app) | | tool call to connector v MCP CLIENT ▶▶▶ REMOTE SERVER ▶ GOOGLE DRIVE (connector — (hosted by (external your machine) Anthropic) service) Streamable HTTP HTTP request (over internet) to API

The connector (client) is still on your machine. The server (program) is remote — on Anthropic's servers. This is what you used last week.

The Key Difference

	Claude Code (local)	Claude Desktop (remote)
Client (connector)	On your machine	On your machine
Server (program)	On your machine	On Anthropic's servers
How they talk	stdio (local programs talking)	Streamable HTTP (over internet)
You build it?	Yes	No — Anthropic builds it

The connector is always on your machine. The difference is where the program runs.

Component Definitions

MCP Term	What It Is	Where It Runs
Host	The app you type into	Your computer
Client	Connector — one per server, automatic	Your computer
Server (local)	Program you build — calls an API	Your computer
Server (remote)	Same job, Anthropic hosts it	Anthropic's servers
External service	Google Drive, Slack, Zoom etc.	The internet
Claude model	Decides which tool to call. Not part of MCP.	Anthropic's servers

Tools — The Part That Matters

MCP defines three primitives: Tools, Resources, and Prompts. Only Tools matter in practice.
Tools = actions the server can perform (search, create, update, delete)
Model-controlled: Claude decides which tool to call based on your question. You just ask in plain language.
Tool discovery is automatic — client calls tools/list on the server at startup

The Request-Response Flow

1

You type: "Search my Google Drive for Q4 reports"

2

Host (Claude Code) sends your message to the Claude model

3

Model decides to call search_files tool

4

Host hands the call to MCP client (connector)

5

Client sends call to MCP server (program) via JSON-RPC over stdio

6

Server calls Google Drive API → gets results → formats → returns

7

Client passes result to host → model writes your answer

Steps 4–7 are MCP. You just asked a question. MCP handled the plumbing.

What is CLI?

CLI = Command-Line Interface. Typing text commands instead of clicking buttons.
Terminal = the app where you type (Terminal on Mac, PowerShell on Windows). No buttons, no menus.
Common tools: curl (fetches data from URLs), git, ffmpeg, python
CLI tools run locally, but commands may send data over the internet (curl fetches from a URL, git push sends to GitHub)

How Claude Code Uses CLI

Claude Code has a Bash tool that sends commands to your terminal
The terminal executes the command — same as if you typed it yourself
Claude constructs the command programmatically, then reads the text output
No setup. No config. If the tool is installed on your machine, Claude can use it.

With CLI, you see the exact command and where data is going. With MCP, the server program handles it internally.

CLI vs MCP

Aspect	CLI	MCP
Setup	Install the tool, done	Build server, configure JSON, restart host
Discovery	Claude knows common tools	`tools/list` at startup — automatic
Data format	Raw text/JSON	Structured, typed responses
State	Stateless — each command independent	Persistent connection
Auth	Per command or env vars	Configured once
Ecosystem	Thousands installed already	Growing, still early

MCP vs CLI: When to Use Which

▶ One-off task? (convert a file, check something) CLI

▶ Repeated integration? (CRM, calendar, email) MCP

▶ Simple utility? (git, ffmpeg, file ops) CLI

▶ Complex API with auth? (Google, Slack, Notion) MCP

▶ Need auto-discovery of capabilities? MCP

▶ Just run and read the output? CLI

Claude decides which to use. MCP preferred when available. CLI is the fallback. You don't specify.

Real-World Examples

Task	Approach	Why
"Convert this video to MP3"	CLI (`ffmpeg`)	One command, done
"Search Google Drive for Q4 reports"	MCP	OAuth, pagination, structured
"Create a Jira ticket"	MCP	Auth, structured, repeatable
"Count lines in this file"	CLI (`wc -l`)	Trivial utility
"Pull my Slack messages"	MCP	OAuth, channels, threading
"Run my test suite"	CLI (`npm test`)	Direct command
"Log interaction in CRM"	MCP	Structured data, auth

Live Build: Weather MCP Server

OpenWeatherMap — free tier, API key auth, no OAuth

1

"I want Claude to check the weather whenever I ask"

2

Paste the build prompt into Claude Code

3

Watch it build — files, dependencies, tools

4

Configure — add to config JSON

5

Test with MCP Inspector, then test in Claude Code

Same Task via CLI

Ask Claude Code: "What's the weather in Dublin?"

curl "https://api.openweathermap.org/data/2.5/weather?q=Dublin&appid=KEY&units=metric"

Raw JSON back. Claude reads it, works out which fields matter, writes a response.

	MCP	CLI
Setup	~5 min	0
First use	Slower	Faster
Tenth use	Instant (auto-discovered)	Same as first
Error handling	Structured	Raw HTTP

MCP Security Risks

1. Data Leaks & PII Exposure

Server passes data to external APIs. Sensitive data can leak without you knowing.

2. Prompt Injection

Malicious content from an API could manipulate what Claude does next.

3. Shadow MCP (Unapproved Servers)

Anyone could install a server connecting Claude to unsanctioned services.

4. No Visibility

Without logging, no audit trail of what data went where.

5. Tool Overload

Too many servers overwhelm Claude — wrong tool calls, confusion, errors.

Source: MCP Manager

Security Checklist

1

Review source code before installing any third-party server

2

Environment variables for credentials — never hardcode API keys

3

Minimum scope API keys — read-only if you only need read

4

Build your own when possible — you control every line

5

Test with non-sensitive data — don't connect production CRM on day one

6

Keep server count manageable — connect what you use, disconnect what you don't

Next Steps

Work through the setup checklist — build the server from today's demo
Identify one repeated task in your workflow → build an MCP server for it
CLI for everything else — they complement each other
Post questions in the community

GenAI Skills Academy

skool.com/genai-skills-academy-1964