How to connect your first MCP server to Claude Code or Cursor

If you have spent any time with an AI coding assistant, you have probably hit the moment where it can write code beautifully but has no idea what is in your database, your issue tracker, or your design files. You end up copy-pasting context into the chat over and over. The Model Context Protocol (MCP) exists to kill that copy-paste loop. This guide explains what MCP is and walks you through connecting your first server.

What MCP actually is

MCP is an open standard, introduced by Anthropic, for connecting AI tools to outside data and services. The common analogy is that MCP is a "USB-C port for AI": instead of every tool inventing its own custom integration, an AI agent can plug into any MCP-compatible server using one universal protocol.

The big practical win is portability. Because Cursor and Claude Code both speak the same protocol, a server you configure in one generally works in the other — and as of WWDC 2026, even Apple's Xcode 27 added MCP plugin support. Learn it once, reuse it everywhere.

An MCP server is just a small program that exposes tools (actions the AI can take) and resources (data it can read). There are pre-built servers for GitHub, PostgreSQL, Slack, Google Drive, Puppeteer, and many more.

The two kinds of servers you will meet

There are two transports worth knowing as a beginner:

• Local (stdio) servers run as a process on your own machine — great for things like a filesystem or a local database. They start with a command such as npx or python.
• Remote (HTTP/SSE) servers are hosted by a vendor at a URL, often behind OAuth login — for example Notion, Asana, or Sentry.

Connecting a server in Claude Code

Claude Code uses a single claude mcp add command. For a remote server, point it at a URL:

claude mcp add --transport http notion https://mcp.notion.com/mcp

For a local server, everything after -- is the command Claude runs to start it:

claude mcp add --env AIRTABLE_API_KEY=YOUR_KEY --transport stdio airtable -- npx -y airtable-mcp-server

The -- separator is important: it tells Claude Code where its own flags stop and the server's launch command begins. You can then list what is connected with claude mcp list to confirm the handshake succeeded.

Claude Code also supports scopes, which decide who can see the server. A local scope keeps it to you on this machine, project scope writes the config into a shared .mcp.json file your whole team can commit to git, and user scope makes it available across all your projects. For anything your teammates also need, project scope is usually the right call.

Connecting a server in Cursor

Cursor exposes the same capability through its UI. Open Cursor Settings, go to Features > MCP, and click "+ Add New MCP Server", then paste either the launch command (for local servers) or the URL (for remote ones). Under the hood Cursor stores this in a JSON config, so you can also edit it by hand. A nice 2026 refinement: Cursor's agent now loads tool descriptions on demand rather than stuffing every server's tools into the prompt, which keeps things fast even when you have several servers connected.

A concrete first project

If you want a five-minute win, connect a GitHub MCP server and then ask your assistant something it could never answer before:

"Look at the three most recently closed issues in this repo and summarize what changed."

Without MCP, the assistant would guess or ask you to paste the issues. With the server connected, it calls the GitHub tools directly, reads the real issues, and answers from live data. That is the whole point of MCP in one example: the model stops being a clever text box and starts being an agent that can reach into the systems you actually work in.

Safety notes before you go wild

MCP gives an AI real access to real systems, so treat new servers like any other dependency. Prefer official or well-reviewed servers, read what tools a server exposes before connecting it, scope API keys to the minimum permissions needed, and be cautious with servers that can write or delete data. Start with read-only use cases — querying your database schema, reading issues, searching docs — and graduate to write actions once you trust the setup.

Once the copy-paste loop is gone, it is hard to go back. Start with one server that solves a real annoyance in your workflow, and grow from there.