Claude Code is very good with the files in a folder on your computer. Your business mostly doesn't live there. It lives in Notion, Stripe, your email platform and your issue tracker. MCP is how you let Claude reach those places directly, instead of you copying and pasting between tabs.
What MCP is
MCP stands for Model Context Protocol. It's an open standard for connecting AI apps to other software. An MCP server is a small adapter that a tool's maker (or anyone) publishes. It tells Claude what it can do in that tool, such as "search pages", "list customers" or "create a draft", and does it when asked. Claude Code is the client that connects to it.
Once a server is connected, those abilities show up as tools Claude can use, right next to reading files and running commands. Anthropic's own rule of thumb for when to bother: connect a server when you notice you keep copying data into the chat from another tool.
Local and remote servers
Servers come in two flavors:
- Remote servers run on the vendor's computers. You connect with a URL and sign in through your browser. Anthropic recommends the HTTP transport for these, and most official business servers work this way.
- Local servers are programs that run on your machine. Claude Code starts them for you, often through
npx, and you usually hand them an API key. They run with your user account's full access.
You may also see SSE servers in older guides. That transport is deprecated; use HTTP where the vendor offers it.
Adding a server
One command, run in your normal terminal (not inside a Claude Code session):
claude mcp add --transport http notion https://mcp.notion.com/mcp
That's the transport, a name you choose (notion), and the server's URL. The name matters later: Claude Code labels the server's tools mcp__notion__..., and that's what you'll use in permission rules.
A local server looks slightly different, because Claude Code needs the command to start it. Everything after -- is the server's own command:
claude mcp add --env AIRTABLE_API_KEY=YOUR_KEY --transport stdio airtable -- npx -y airtable-mcp-server
To look after your servers:
claude mcp list # every server and whether it's connected
claude mcp get notion # details for one server
claude mcp remove notion # remove it
Inside a session, type /mcp to see each server's status and how many tools it offers.
Scopes: who gets the server
The --scope flag decides where the server is available and who else gets it.
| Scope | Loads in | Shared? | Stored in |
|---|---|---|---|
local (the default) | This project, just you | No | ~/.claude.json |
project | This project, for everyone | Yes, through version control | .mcp.json in the project folder |
user | All your projects | No | ~/.claude.json |
A simple way to choose: user for tools you use everywhere, like your notes app. project when a team shares a repo and should share the setup; each person still signs in with their own account. local for experiments, or for anything that needs a key you don't want in a shared file. Claude Code asks before it uses servers from a project's .mcp.json, so a downloaded repo can't connect itself to things silently.
Signing in
Most remote servers use OAuth, the "Sign in with..." flow you already know. After adding the server, start Claude Code, type /mcp, pick the server and follow the browser login. Claude Code stores and refreshes the token. If a login expires, /mcp has a re-authenticate option, and "Clear authentication" revokes Claude Code's access.
You can also sign in from your shell with claude mcp login notion. That matters for scheduled or scripted runs, which can't open a browser, so sign in interactively first. More on that in running Claude Code on a schedule.
If you use Claude on the web, there's a shortcut: connectors you've added at claude.ai appear in Claude Code automatically when you're logged in with the same account.
Five official servers worth knowing
Each of these comes from the vendor's own documentation. Check the linked page before you connect, because URLs and plans change.
Notion
claude mcp add --transport http notion https://mcp.notion.com/mcp
Then run /mcp and sign in. Try: "Find every page in my content calendar due this month and draft an outline for each in drafts/."
Stripe
claude mcp add --transport http stripe https://mcp.stripe.com/
Sign in with /mcp. Stripe's consent screen lets you grant access to live accounts or sandboxes with different permissions for each, so start with a sandbox. Its tools include stripe_api_read and stripe_api_write, and Stripe asks for human confirmation, through a link, before certain writes such as refunds. If you'd used API keys before: Stripe says that from December 2026 its MCP server stops accepting full-access secret keys and restricted keys without the Agent tag, so use OAuth or an agent API key.
Try: "Summarize last month's failed payments by reason. Don't change anything."
Supabase
claude mcp add --scope project --transport http supabase "https://mcp.supabase.com/mcp?project_ref=abc123&read_only=true"
Replace abc123 with your project's ID. The two URL settings are Supabase's own safety switches: project_ref limits the server to one project, and read_only=true runs every query as a read-only database user. Supabase's advice is to connect a production project only when the task really needs production data.
Linear
claude mcp add --transport http linear-server https://mcp.linear.app/mcp
Sign in with /mcp. Try: "Turn the feedback in research/interviews.md into Linear issues, but show me the list before you create anything."
Kit
Kit (the email platform formerly called ConvertKit) runs an official server at https://app.kit.com/mcp, available on its paid Creator and Creator Pro plans and signed in with OAuth. Kit's docs give the URL rather than a Claude Code command, so use the standard one:
claude mcp add --transport http kit https://app.kit.com/mcp
It can read and change subscribers, tags, sequences and broadcasts, including sending them. Try: "Pull my last five broadcasts and their stats, then draft three subject lines in the style of the best one." Keep anything that sends on a prompt, and only email people who opted in (the usual CAN-SPAM and GDPR rules still apply when an AI pressed the button).
Security: only plug in what you trust
An MCP server can read your data and act as you. Treat adding one like installing an app with your passwords.
- Use the vendor's official server. Get the URL from the vendor's own docs or Anthropic's connector directory. Anthropic reviews directory listings against its criteria but says it doesn't security-audit any MCP server. Be wary of unofficial packages promising your favorite tool.
- Look at what it exposes. After connecting, ask: "List every tool from the stripe server and mark which ones change data." The
/mcppanel shows the tool count per server. - Watch for prompt injection. That's when text Claude reads (a web page, an inbound email, a support ticket) contains instructions aimed at Claude. Be extra careful when one server reads outside content and another can send money or email.
- Gate the write tools. MCP tools follow the same permission rules as everything else. Allow the safe ones by name and keep risky ones on ask, which prompts even in auto mode:
{
"permissions": {
"allow": ["mcp__stripe__stripe_api_read"],
"ask": ["mcp__stripe__stripe_api_write"]
}
}
- Remember where local servers run. A local server is a program on your computer with your access, and it runs outside Claude Code's sandbox.
- Clean up. Remove servers you no longer use. For OAuth servers, run
claude mcp logout <name>beforeclaude mcp remove <name>, so the token is revoked at the provider where it supports that, not just deleted on your machine.
More on rules in stop clicking Allow all day, and on limits for unattended runs in keeping agents cheap and safe.
Start with one
Pick the tool you copy from most. Add its official server at user scope, sign in with /mcp, and give Claude one read-only job. Once you trust the results, add the next one. Five well-understood servers beat twenty you've never looked at.