How to use MCP with Claude Code

Out of the box, Claude Code is brilliant inside one boundary: your files and your terminal. It can read your project, edit it, run commands, and reason about all of it. But the moment a task needs something outside that boundary — “what does the failing ticket say?”, “pull last week’s numbers from the database”, “check whether the live site actually loads” — you’re back to copying and pasting.

MCP is what erases that boundary.

What MCP actually is

MCP — the Model Context Protocol — is a standard way for AI assistants to talk to external tools and data. Think of it as a universal adapter. Instead of every AI app inventing its own way to connect to, say, your issue tracker, they all speak one shared language. A tool exposes an MCP server; Claude is an MCP client; they connect, and now Claude can use that tool.

The “open standard” part matters. Because MCP is shared, a server someone built for one app works in any app that speaks MCP. The ecosystem of connectors grows for everyone at once.

A useful mental model: MCP servers give Claude new senses and new hands. A browser server lets it see and click web pages. A database server lets it query your data. A docs server lets it read your knowledge base. Each one extends what “look into this and handle it” can mean.

Why it’s the big unlock

Here’s the same task, before and after.

Before MCP: You open the ticket, copy the description, paste it into Claude, copy the error log, paste that too, describe your database schema from memory, and then ask for a fix.

After MCP: “Read ticket #4012, find the cause in the code, check the relevant rows in the database, and propose a fix.”

That’s not a small convenience. It collapses a multi-tool, copy-paste-heavy chore into one sentence — and it removes the transcription errors that creep in every time you move information by hand.

Setting up your first server

The exact command depends on the server, but the shape is always the same. Most servers are added with a single command:

claude mcp add <name>

Or, for servers you configure by hand, you add a small block to your settings describing how to start the server and what it’s allowed to do. Once it’s connected, Claude lists the new capabilities and can start using them — asking your permission before it takes any real action.

A good first connection is something low-stakes and obviously useful:

• A browser server (like Playwright) so Claude can open pages, take screenshots, and verify your site works.
• A docs or notes server (like Notion) so it can read and search your knowledge base.
• A calendar or email server so it can draft, schedule, and look things up.

Start with one. Get comfortable with how the permission prompts feel. Then add the next.

The permission model — and why it matters

Connecting a tool means granting access, so treat each new MCP server the way you’d treat installing a new app:

• Connect servers you trust. A server runs with whatever access you give it.
• Scope it down. Many servers let you limit what they can touch — read-only where possible, a single project rather than everything.
• Watch the prompts. Claude asks before it does anything consequential through a server. Those prompts are your safety rail; read them rather than reflexively approving.

This is the same discipline as the rest of Claude Code’s permission system: the model proposes, you approve, and irreversible actions get a closer look. MCP doesn’t change that contract — it just widens the set of things Claude can propose.

When to reach for MCP (and when not to)

Reach for it when a task repeatedly forces you to leave Claude to fetch something — the same ticket system, the same dashboard, the same docs, every day. That repetition is the signal that a connection will pay off.

Don’t bother when the thing you need is already a file in your project, or when it’s a one-off you’ll never repeat. The whole point of MCP is to automate a recurring bridge between Claude and a tool you live in. For a single lookup, pasting is faster than configuring.

Where this leads

Once Claude can see your tools, the nature of what you delegate changes. You stop asking it to “help me write the code that queries the database” and start asking it to “find out why orders dropped on Tuesday” — and it goes and looks. That’s the same leap that turns an assistant into an agent: the ability to take actions in the world, observe the result, and decide what to do next.

MCP is the plumbing that makes that leap possible. Set up one server this week. The first time you ask a question that spans your code, your data, and the web — and get one coherent answer back — you’ll understand why people who use it never go back.