In Claude Code, MCP (the Model Context Protocol) is an open standard that lets Claude connect to your outside tools and data: things like GitHub, Slack, Jira, Google Drive, and a Postgres database. Once you connect one of these systems as an MCP server, Claude can read from it and act on it directly, instead of you copying information back and forth into the chat. If you have ever found yourself pasting an issue, a spreadsheet, or a customer record into Claude just so it can see it, MCP is what removes that step.

What the Model Context Protocol is

The Model Context Protocol is a shared standard for connecting an assistant like Claude to external tools, APIs, databases, and data sources. "Open standard" just means it isn't owned by any one company, so any tool can build a connection to it and Claude Code can talk to all of them the same way. Each connection is called an MCP server. A server is a small piece of software that sits between Claude and one of your systems and hands Claude a set of actions it's allowed to take there.

Practically, that turns requests like these into something Claude can just do:

  • Open a pull request on GitHub from a change you described.
  • Query your Postgres database to pull a list of customers.
  • Post an update to a Slack channel, or read what a design says in Figma.
  • Pick up a task straight from a Jira issue and act on it.
connected servers
GitHubRead issues, open pull requests, review code
SlackPost messages, read a channel's history
Postgres databaseQuery records, look up a schema
JiraPull an issue, act on the work it describes

A rough picture of connected servers. The last one is configured but switched off.

How you connect an MCP server

You add a server from your terminal with the claude mcp add command, and you manage the connections from inside Claude Code with the /mcp command. There are three ways a server can run, and the difference is mostly about where the server lives.

Option 1Remote (HTTP)The recommended way for cloud services. The server runs somewhere online and Claude connects over the web.
Option 2Local (stdio)The server runs as a program on your own machine. Good for tools that need direct access to your computer.
LegacySSEAn older remote type that is now deprecated. Use HTTP instead where you can.

Many remote services just need you to sign in. The /mcp command handles that sign-in for you (the technical name is OAuth), and it's also where you reconnect a server, check which servers are connected, and enable or disable them.

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

Connecting a remote server, using Notion as the example. Click to copy.

Where the connection is saved

Once you add a server, the setup is saved to a small configuration file so it's there next time. You mostly won't touch this file by hand, but it helps to know it exists. A server saved to a project's .mcp.json file can be checked into version control, which means everyone on your team gets the same connections. A server saved to your own user settings is private to you and follows you across every project.

Good to know: connecting a server is about permission, not surrender. Each server hands Claude a defined set of actions, and you stay in control of which servers are switched on. If you don't trust a source, don't connect it.

A real example

Say a teammate files a bug in Jira and drops a note about it in Slack. Normally you'd read the Jira ticket, copy the details into Claude, make the fix, then go open a pull request in GitHub yourself. With those three systems connected as MCP servers, you can hand the whole chain to Claude in one request:

claude
› Fix the bug described in Jira issue ENG-412 and open a PR on GitHub # reading Jira issue ENG-412... # found the failing case, making the change... # opening pull request on GitHub... done PR opened, linked back to ENG-412

Claude read the ticket straight from Jira, made the change, and opened the pull request, all without you copying anything between windows. The same pattern works for pulling live numbers from a database into a report, or checking a design in Figma before writing the copy that goes with it.

Why you can connect a lot of tools

You might worry that hooking up ten tools would overwhelm Claude with options. It doesn't, thanks to a feature called tool search. Instead of loading every server's full instructions up front, Claude Code loads only the tool names at the start and pulls in the full details for a tool only when a task actually needs it. In plain terms: you can connect many servers without slowing Claude down or crowding out the conversation. Connect the tools you use and let Claude reach for the right one when the moment calls for it.