What Are MCPs? Your First MCP Setup in Claude Code

MCPs let Claude Code talk to external tools directly. Notion, databases, GitHub, Slack, whatever you need. This guide explains what they are from scratch and walks you through setting up your first one: the Notion MCP.

What is an MCP?

MCP stands for Model Context Protocol. It's a standard that lets AI models connect to external tools and data sources.

Think of it like a USB port. Your laptop is useful on its own. Plug in a monitor, a drive, a keyboard and it becomes a workstation. MCP does the same thing for Claude. Without MCPs, Claude can only work with what's in your conversation and your local files. With MCPs, Claude can read your Notion workspace, query your database, create GitHub issues, send Slack messages. All from one prompt.

The key thing: Claude doesn't just get data from these tools. It can write back. Ask Claude to update a Notion page and it updates the page. Ask it to create a GitHub PR and it creates the PR.

How MCPs actually work

Three components:

MCP Server. A small program that talks to one specific tool. The Notion MCP server reads and writes Notion pages. The Postgres one runs SQL queries. One server per tool.

MCP Client. The AI tool that connects to MCP servers. Claude Code is an MCP client. So is Cursor, Windsurf, and most AI coding tools now.

The Protocol.The standard between them. Defines how the client finds out what the server can do, sends requests, gets responses back. You don't need to understand the protocol to use MCPs. Just know it's the reason any server works with any client.

In practice: you point Claude Code at an MCP server. Claude connects on startup, figures out what tools the server offers, and makes them available. Ask Claude to do something that needs that tool and it calls the server automatically.

What can you actually do with MCPs?

Some real examples:

Notion MCP. Claude reads and writes your Notion workspace. Search pages, update databases, add new entries. Good for content calendars, project tracking, knowledge bases.
Postgres MCP. Claude runs read-only SQL against your database. Ask "how many users signed up this week" and get the real number back.
GitHub MCP. Create issues, open PRs, read code from repos you don't have cloned locally.
Slack MCP. Read channels, send messages, pull thread summaries.
Filesystem MCP. Read and write files outside the current project directory. Good for cross-project workflows.

There are hundreds of MCP servers. Anthropic maintains official ones for the major tools. The community builds the rest. Quality varies, so start with official servers.

Step 1: find an MCP server

For this guide, we're setting up the Notion MCP. It's official, well-maintained, and immediately useful.

The server package is @anthropic-ai/notion-mcp-serveron npm. You don't need to install it manually. Claude Code handles that.

You'll need a Notion integration token:

Go to notion.so/my-integrations
Click "New integration"
Name it whatever you want (e.g., "Claude Code")
Copy the integration token (starts with ntn_)
Go to the Notion pages you want Claude to access and share them with your integration (click "...", Connections, add your integration)

That last step matters. The integration can only see pages you explicitly share with it.

Step 2: add the MCP to Claude Code

Open your Claude Code config. Run this in your terminal:

claude mcp add notion -e OPENAPI_MCP_HEADERS='{"Authorization":"Bearer ntn_YOUR_TOKEN_HERE","Notion-Version":"2022-06-28"}' -- npx -y @anthropic-ai/notion-mcp-server

Replace ntn_YOUR_TOKEN_HERE with your actual token.

That's it. One command. Claude Code now knows about the Notion MCP server and will start it automatically.

Step 3: test it

Start a new Claude Code session:

claude

Ask Claude something that requires Notion access:

Search my Notion workspace for pages about "content ideas"

Claude calls the Notion MCP server, searches your workspace, and returns the results. If it works, you're connected.

Try a write operation:

Add a new entry to my content ideas page: "MCP explainer video - May 2026"

Claude reads the page structure, figures out the format, adds the entry.

Step 4: understanding the config

Running claude mcp add added an entry to your Claude Code MCP config. You can see it with:

claude mcp list

Under the hood, the config tells Claude Code:

What command to run to start the MCP server (npx -y @anthropic-ai/notion-mcp-server)
What environment variables to pass (your auth token and the Notion API version)

When Claude Code starts, it launches the MCP server as a subprocess and connects via the protocol. The Notion server exposes tools like search, read_page, update_page, create_page.

Adding more MCPs

Same pattern for any MCP. Find the server, get auth credentials, run claude mcp add. Some examples:

GitHub:

claude mcp add github -e GITHUB_PERSONAL_ACCESS_TOKEN=ghp_YOUR_TOKEN -- npx -y @modelcontextprotocol/server-github

Postgres (read-only):

claude mcp add postgres -- npx -y @modelcontextprotocol/server-postgres postgresql://user:pass@localhost:5432/mydb

Each MCP server has its own auth setup. Check the server's README for specifics.

The gotcha: when MCPs are worth it vs overkill

MCPs add a connection that Claude manages in the background. For tools you use daily alongside code, they cut out a lot of tab-switching and copy-pasting. For tools you touch once a month, the setup isn't worth it.

Worth it when:

You reference the tool multiple times per session (Notion for planning, GitHub for issues)
The tool has data Claude needs to make decisions (database queries, analytics)
You want Claude to take actions, not just read (create issues, update pages)

Overkill when:

You use the tool once in a blue moon
The tool's MCP server is community-built and poorly maintained
You just need to paste in a single value (faster to copy-paste than to set up an MCP)

Start with one MCP. Get used to the workflow. Add more when you feel the friction of not having them.