How Do You Add an MCP Server to Claude Code?
You add a remote MCP server to Claude Code with one claude mcp add --transport http line, though some vendors now ship a plugin instead. MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems (modelcontextprotocol.io). Claude Code MCP setup comes in three forms, per Anthropic's Claude Code MCP docs:
- HTTP, for remote servers:
claude mcp add --transport http <name> <url>. The docs call HTTP the recommended option. The older SSE transport is deprecated. - stdio, for local servers:
claude mcp add [options] <name> -- <command> [args...]. The--separates Claude's own flags, like--scopeand--env, from the command that starts the server. - Plugin:
/plugin install <plugin>@<marketplace>inside a session, orclaude plugin installin the terminal where the vendor documents it (Figma and Exa do). A plugin bundles the server, so you add or remove it by installing or uninstalling the plugin.
Every server also lives in a scope. Pass --scope (or -s) to choose:
| Scope | Who sees it | Stored in |
|---|---|---|
| local (default) | you, this project | your user config file (~/.claude.json) |
| project | your team, via git | a config file in the project root (.mcp.json) |
| user | you, every project | your user config file (~/.claude.json) |
My rule of thumb: use local to try a server, project for team tools you want in git, and user for something you want everywhere, like a docs server. Never commit a token in .mcp.json. Claude Code asks for your approval in interactive sessions before it uses project-scoped servers from that file.
After adding a server that signs in with OAuth, open Claude Code and run /mcp. It shows each server's status and runs the sign-in. From the terminal, claude mcp list shows everything you've configured and claude mcp remove <name> takes a server out. The same steps work for every one of the Claude Code MCP servers below.
Which Claude Code MCP Servers Are Best for Coding?
The best Claude Code MCP servers for coding are GitHub and Context7, plus a browser server for UI tests. Each Claude Code MCP server in this group is run by its own vendor. Add Sentry once you have production errors to chase.
GitHub MCP Server: Repos, Issues and PRs
GitHub's official server connects Claude Code to your repos, issues and pull requests. For Claude Code 2.1.1 and newer, GitHub's install guide gives this, to run in your normal terminal:
claude mcp add-json github '{"type":"http","url":"https://api.githubcopilot.com/mcp","headers":{"Authorization":"Bearer YOUR_GITHUB_PAT"}}'
On 2.1.0 and older, the guide uses the legacy form:
claude mcp add --transport http github https://api.githubcopilot.com/mcp -H "Authorization: Bearer YOUR_GITHUB_PAT"
- Access (as of November 2026): a GitHub account and a personal access token; MIT licence; the guide names no price.
- Best for: issue and PR work without leaving the terminal.
- Limit: it needs a token. GitHub suggests keeping it in a
.envfile and adding.envand.mcp.jsonto.gitignore.
Context7: Current Library Docs
Claude's training data lags behind library releases. Context7 gives Claude Code current library documentation instead. Its Claude Code page leads with:
npx ctx7 setup --claude
Or install the plugin inside a session:
/plugin marketplace add upstash/context7
/plugin install context7@context7-marketplace
- Access (as of November 2026): the setup command creates an API key through OAuth; the plugin needs no key. The page doesn't state a price.
- Best for: coding against fast-moving frameworks.
- Limit: the page shows no
claude mcp addform, so a manual add means following Context7's general client instructions.
Playwright MCP: Browser Tests
The Playwright MCP server from Microsoft drives a real browser through Playwright. It reads pages as structured accessibility snapshots, so it doesn't need screenshots or a vision model. The README gives:
claude mcp add playwright npx @playwright/mcp@latest
- Access (as of November 2026): open source under Apache-2.0, runs locally.
- Best for: UI checks and scripted test flows.
- Limit: it needs Node.js 20 or newer.
Chrome DevTools MCP: Speed and Debugging
Chrome DevTools MCP is the pick when the problem is speed rather than function. It records performance traces in a live Chrome, reads network requests and console messages with source-mapped stack traces, and takes screenshots. The repo documents a plugin:
/plugin marketplace add ChromeDevTools/chrome-devtools-mcp
/plugin install chrome-devtools-mcp@chrome-devtools-plugins
- Access (as of November 2026): open source under Apache-2.0; needs Node.js LTS, current Chrome and npm.
- Best for: performance traces and rendering bugs.
- Limit: it collects usage statistics by default (opt out with
--no-usage-statistics), and it exposes the browser's content to Claude, so keep sensitive logins out of that browser.
Sentry MCP: Production Errors
Sentry's server lets Claude read your errors, issues, traces and performance data. The repo gives a CLI route; run sentry auth login first, and the config needs no token:
claude mcp add sentry -- sentry mcp
There's also a plugin that adds a sentry-mcp subagent:
claude plugin marketplace add getsentry/sentry-mcp
claude plugin install sentry-mcp@sentry-mcp
- Access (as of November 2026): a Sentry account; no price stated for the server.
- Best for: debugging production errors from the terminal.
- Limit: the AI search tools, such as
search_errorsandsearch_issues, need an LLM provider key. Without one they're unavailable; the other tools still work.
The next group connects Claude to the places where your team plans, designs and stores data.
Best Claude Code MCP Servers for Product, Design and Data
These Claude Code MCP servers pay off only for tools your team already works in. A Jira server does nothing for you if your issues live in Linear. I use two of these myself: Figma's server for design files and Notion's for docs.
Linear MCP: Issues From the Terminal
Linear's server finds, creates and updates issues, projects and comments. From Linear's docs:
claude mcp add --transport http linear-server https://mcp.linear.app/mcp
Then run /mcp in a session to sign in.
- Access (as of November 2026): a Linear account, OAuth 2.1 or an API key; the docs list no plan requirement for the server.
- Best for: teams that track work in Linear.
- Limit: it's read-write by default. For read-only access, use
https://mcp.linear.app/mcp/readonlyas the URL.
Notion MCP: Docs and Briefs
I use Notion's server for docs and briefs, and it's the one I'd add first as a marketer. It can search, read and update the pages you can access. From Notion's setup page:
claude mcp add --transport http notion https://mcp.notion.com/mcp
Then run /mcp and finish the OAuth flow.
- Access (as of November 2026): works on every Notion plan; MCP tokens and higher request limits are on Business and Enterprise.
- Best for: specs, briefs and wikis.
- Limit: guests can't connect. Only workspace members can.
Figma MCP: Design to Code
The Figma MCP server brings design context into Claude Code. I use it for design files: paste a link to a frame or layer into your prompt and Claude works from it. Figma's install page recommends the plugin, which also bundles Figma's skills:
claude plugin install figma@claude-plugins-official
Or add the server by hand:
claude mcp add --transport http figma https://mcp.figma.com/mcp
- Access (as of November 2026): available on all Figma plans with a supported client; you sign in with your Figma account on first connect.
- Best for: turning designs into code.
- Limit: your organization may block plugin installs. Then use the manual command, which comes without the skills.
Atlassian MCP: Jira and Confluence
Atlassian's server connects Claude to Jira, Jira Service Management, Confluence, Bitbucket and Loom. From Atlassian's guide:
claude mcp add --transport http atlassian https://mcp.atlassian.com/v2/mcp
Then run /mcp to sign in.
- Access (as of November 2026): OAuth 2.1; no plan tier named; enriched Teamwork Graph calls use Rovo credits.
- Best for: teams on Jira and Confluence.
- Limit: v1 usage moves to v2 tools on March 1, 2027, so use the v2 URL above.
Supabase MCP: Postgres Backends
The Supabase MCP server lets Claude query and manage your Supabase projects. Supabase's docs add it at project scope:
claude mcp add --scope project --transport http supabase "https://mcp.supabase.com/mcp"
- Access (as of November 2026): a Supabase account with browser sign-in, no personal access token; creating a project or branch can trigger a cost confirmation.
- Best for: apps built on Supabase Postgres.
- Limit: Supabase names prompt injection as the main risk: a tricked model could follow untrusted commands and expose data. Use read-only mode by adding
?read_only=trueto the URL, andproject_ref=<id>to limit it to one project.
Stripe MCP: Billing and Payments
Stripe's server gives Claude access to the Stripe API and a search over Stripe's docs. From Stripe's MCP page:
claude mcp add --transport http stripe https://mcp.stripe.com/
Then run /mcp to sign in with OAuth. Stripe also offers a plugin, set up through its CLI with stripe agent setup.
- Access (as of November 2026): a Stripe account, OAuth or an agent API key; some actions, like creating a refund, need your approval in the Dashboard.
- Best for: billing, subscriptions and payment debugging.
- Limit: from January 11, 2027, Stripe's server stops accepting full-access secret keys. OAuth sessions expire after 14 days.
The last two picks deal with the open web.
Best Claude Code MCP Servers for Web Research and SEO
For research you need only two Claude Code MCP servers: one for search, plus one free check before you ship a page. If SEO is your main job, my best SEO MCP servers guide goes further.
Exa MCP: AI Web Search
The Exa MCP server gives Claude web search, page fetching and a multi-step research agent. Exa's docs give one Claude Code command:
claude plugin install exa@claude-plugins-official
- Access (as of November 2026): search and fetch work without a key at a free rate limit; the Exa Agent tool needs OAuth or an API key.
- Best for: research tasks where Claude needs fresh pages.
- Limit: the free rate limit isn't published. A 429 error means you need to sign in or add a key.
AI Crawlability Check: AI Crawler Access
I built this one, so weigh my pick accordingly. AI Crawlability Check is a free MCP server that checks whether 29 AI crawlers, including GPTBot, ClaudeBot and PerplexityBot, can reach a page, whether the page can be indexed, and how quotable it is. 70 measured checks across four tools, no account or API key. It's listed in the official MCP Registry, Glama, mcpservers.org and the GitHub MCP Registry.
claude mcp add --transport http ai-crawlability https://mcp.milan-novotny.com/mcp
- Access (as of November 2026): free, open source, no account.
- Best for: checking a page before you publish it.
- Limit: it does not check whether a page is cited today.
You can see the five tools on its page. Whichever servers you add, check who runs them first.
Are Third-Party MCP Servers Safe in Claude Code?
Third-party MCP servers are safe enough in Claude Code if you use the vendor's own server and the narrowest scope that does the job. MCP server security comes down to who wrote the server and what it can touch. Anthropic's Claude Code docs say to verify you trust each server; servers that fetch outside content can expose you to prompt injection. I follow three rules for every Claude Code MCP server I add:
- Use the vendor's server. The old reference GitHub, Brave Search, Sentry and Slack servers in the MCP project are now archived. Every pick above is run by the company behind the product.
- Pick the narrowest scope and mode. Local scope until you're sure, then project scope for the team. Use read-only endpoints where they exist, like Linear's
/readonlyURL and Supabase'sread_only=true. - Read the approval prompt. Claude Code asks before it uses project-scoped servers from a repo's
.mcp.json. If you didn't expect a server there, say no and check who added it.
Injected instructions arrive through servers that read the web, like Exa or a browser server, so I keep those away from write access to production data.
FAQ: Claude Code and MCP
Which command lists my servers?
claude mcp list in the terminal; /mcp inside a session. claude mcp get <name> shows the details of one server, and claude mcp remove <name> deletes it.
What does /mcp do?
/mcp shows each server's status and runs sign-in. You use it after adding a remote server that needs OAuth, and you can also turn a server off there without removing it.
Can I share MCP servers with my team?
Yes: project scope saves them in a file you commit to git. For Claude Code MCP servers that file is .mcp.json in the project root, and each teammate approves it on first use. Keep tokens out of it.
Should I use a plugin or claude mcp add?
Follow the vendor's docs; several now prefer a plugin. A plugin can bundle skills and updates with the server, while claude mcp add gives you only the server.