Lesson 3.4: Introduction to MCP

🎯 What You'll Learn: What MCP is, how it connects Claude to external systems, and how to set up your first integration
⏱️ Time Required: 35 minutes

---

The Last Manual Step

Riley's margin report runs automatically every Monday. The audit log captures every file write. But one thing still bothers her.

"Every Sunday night, I log into Prime Broker portal, click Export, wait, download the CSV, move it to the right folder. Ten minutes, but it's the last manual step. What if Claude could just read the data directly instead of waiting for me to export it?"
— Riley Harper

That's what MCP enables.

---

What MCP Is

MCP (Model Context Protocol) is an open standard that lets Claude connect to external systems, databases, APIs, file systems, SaaS tools, through a consistent interface.

Think of it as the difference between:

Giving your assistant a printout of your margin data (what Claude does without MCP)

Giving your assistant read access to the source system directly (what Claude can do with MCP)

MCP servers are translators. Each one sits between Claude and an external system, translating Claude's requests into the API calls, database queries, or file operations that system understands.

Claude Code  ←→  MCP Server  ←→  External System
"Get margin data"  →  API call  →  Prime Broker portal returns data

The MCP ecosystem has grown significantly. There are now servers for filesystem access, GitHub, Slack, PostgreSQL, Google Drive, Notion, Jira, and hundreds of community-built integrations. If your business uses a tool, there's likely an MCP server for it.

---

Where MCP Servers Are Configured

MCP servers are configured in a .mcp.json file in your project root (separate from settings.json, which holds permissions and hooks):

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/riley/margin-automation/data",
        "/Users/riley/margin-automation/outputs"
      ]
    }
  }
}

The args list for the filesystem server is a list of folders Claude is allowed to access. Everything outside those folders remains inaccessible, this is the security model.

Changes to MCP configuration require a restart of Claude Code to take effect.

💡 The easy way (no JSON editing): You don't have to hand-edit this file. The simplest method is the claude mcp add command, or just ask Claude in plain English, "Add the filesystem MCP server scoped to these two folders: [paths]." Claude writes the configuration for you. The JSON above is shown so you understand what gets created and can check it. For a server your whole team should get automatically, ask Claude to add it at "project" scope, which saves it to a shared .mcp.json file you commit with the project.

---

Exercise: Set Up Filesystem MCP (35 min)

Step 1: Understand What You're Installing (5 min)

The filesystem MCP server is the right starting point because:

It requires no accounts or API tokens

It gives you immediate value (Claude can read and write files directly)

It's low risk (you control exactly which folders it can see)

What it enables:

Before MCP:

You: "Analyze my margin data"
Claude: "Please paste the CSV or provide the path and I'll tell you the commands to run"

After MCP:

You: "Analyze the latest export in my data folder"
Claude: [reads the file directly] "I can see 47 open positions, £2.3M net exposure..."

Step 2: Install and Configure (15 min)

Install the filesystem server (requires Node.js):

# Check if Node.js is available
node --version

# If not installed, get it at nodejs.org first

Add to .mcp.json in your project root:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/yourname/your-project/data",
        "/Users/yourname/your-project/outputs"
      ]
    }
  }
}

Replace the paths with folders that actually exist and contain data you want Claude to access. Keep the scope narrow, specific project folders, not your entire home directory.

Riley's configuration:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/riley/margin-automation/data",
        "/Users/riley/margin-automation/outputs"
      ]
    }
  }
}

Step 3: Restart and Test (10 min)

Restart Claude Code after editing .mcp.json.

Check connection status:

/mcp

This shows all configured MCP servers and whether they're connected. The filesystem server should show as active.

Test it:

Ask Claude: "What files are in my data folder?"

If MCP is working, Claude will list them directly. If it fails, it will say it doesn't have access.

Debug if needed:

# Check MCP server status and details
/mcp

Common failures:

Node.js not installed, install it from nodejs.org

Path doesn't exist, verify the folder exists on your machine

Restart not done, MCP config only loads at startup

Step 4: Document Your MCP Setup (5 min)

Add to your CLAUDE.md:

## MCP Configuration

### Active Servers
- **filesystem** — Read/write access to `/data` and `/outputs` folders

### What Claude Can Access
- Prime Broker portal CSV exports (read from `/data`)
- Generated reports (write to `/outputs`)

### What Claude Cannot Access
- Everything outside the configured folders
- Credentials or config files

This is useful when you share the project, whoever else uses it will know immediately what's configured.

---

Finding Other MCP Servers

The filesystem server you just installed is one of Anthropic's actively maintained reference servers (@modelcontextprotocol/server-filesystem). Others in that maintained set include Fetch (web content), Git, and Memory.

Beyond those, there are MCP servers for the tools most businesses use, GitHub, Slack, Google Drive, databases like PostgreSQL, Notion, and many more. These are now maintained by the vendors themselves or the community rather than shipped in one central package, so the install path differs for each.

The reliable way to find the current server for any tool: ask Claude ("Is there an MCP server for Notion, and how do I add it?"), check the tool's own documentation, or browse the official MCP registry. Don't rely on a hardcoded package name from a course or a blog post, MCP integrations move fast, and the maintained source is always more current than any list written months ago.

💡 If you use the Claude desktop app, many of these integrations are available as one-click connectors, no JSON editing required. For a non-technical setup, that's usually the easiest route.

---

✅ Success Criteria:

---

Next Up

Lesson 3.5: Connect to Business Systems

You have filesystem access. Now you'll connect Slack so Claude can post reports directly to a channel, completing the end-to-end automation that runs from data export to notification without you touching it.