Best MCP Servers for Claude Code: 13 Worth Adding

The best MCP servers for Claude Code are the ones that put a system you already use behind a tool call: Sentry for production errors, Context7 for current library docs, Playwright for a real browser, Figma for the designs you're building from, and a hosted database, deploy, or payments server for the systems your app runs on. This list covers 13 of them, with the install command and the one thing to check before you trust each one.

Claude Code adds a remote server with claude mcp add --transport http <name> <url> and a local one with claude mcp add <name> -- <command>. The default scope is local, which stores the server in your user config for the current project only. Pass --scope project to write a .mcp.json your team can commit, or --scope user to enable a server in every project on your machine. Remote servers that use OAuth prompt for sign-in the first time you run /mcp inside a session, and claude mcp list shows which servers are connected. Those four commands are all the setup vocabulary this list needs, and the Claude Code MCP docs cover header-based auth and environment variable expansion if you need them.

How this list was chosen

Every server here is published by the company that owns the underlying system, or is the open-source tool people reach for in its category. Each one had to do a job Claude Code can't already do with its built-in file, shell, and web tools. That rules out generic filesystem and terminal servers, which several ranking posts still include even though Claude Code reads files and runs commands on its own. Where a server offers a read-only or project-scoped mode, that mode is named, because the fastest way to regret an MCP server is to give an agent write access to production on day one. Prices aren't listed because none of these servers charge separately for MCP access; you pay for the underlying product, and Context7 and Playwright cost nothing to run.

1. Context7

Context7 fetches version-specific documentation and code examples for the library you're working with and puts them in the model's context. That matters more than it sounds, since a model's memory of an API is often one or two major versions behind, and the resulting code compiles against a signature that no longer exists. The remote endpoint is https://mcp.context7.com/mcp; a free API key raises the rate limits but isn't required.

claude mcp add --transport http context7 https://mcp.context7.com/mcp

Ask for docs by name in the prompt ("use context7 for the TanStack Router v1 loader API") until you've confirmed the agent reaches for it on its own.

2. Playwright

Microsoft's Playwright MCP drives a real browser through the accessibility tree instead of screenshots, so the agent can fill forms, click through flows, and assert on page state without a vision model. It runs locally over stdio and needs Node.

claude mcp add playwright -- npx @playwright/mcp@latest

It's the most useful server on this list for verifying a UI change end to end, and the reason the "did you actually test it" question stops needing a human at the keyboard.

3. Sentry

Sentry's hosted server at https://mcp.sentry.dev/mcp searches errors, reads stack traces, and triages issues in your projects. Sentry recommends scoping the URL to an organization and project so the agent can't wander:

claude mcp add --transport http sentry https://mcp.sentry.dev/mcp/<org>/<project>

Sign-in is OAuth through /mcp, and once it's connected a "fix the top unresolved issue from the last release" prompt becomes a real workflow rather than a demo, because the agent reads the stack trace itself instead of waiting for you to paste it.

4. Linear

Linear's MCP server at https://mcp.linear.app/mcp finds, creates, and updates issues, projects, and comments, with read-only and read-write access options at sign-in.

claude mcp add --transport http linear https://mcp.linear.app/mcp

The habit it enables is closing the loop: the agent that ships a branch also moves the ticket and writes the comment explaining what changed.

5. Notion

Notion's hosted server at https://mcp.notion.com/mcp searches and edits the pages and databases your account can reach, after an OAuth grant. It's the right server when specs, runbooks, and decision logs live in Notion and you want the agent to read them before it starts, or to write the postmortem after.

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

Notion's MCP documentation lists the supported tools.

6. Stripe

Stripe's server at https://mcp.stripe.com searches the API reference and documentation, and reads or writes Stripe objects through generic stripe_api_read and stripe_api_write tools. Stripe requires human confirmation in the Dashboard for risky writes such as refunds, and you can revoke an agent's OAuth session from your user settings.

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

Prefer OAuth over a pasted restricted key, and start in a sandbox. The Stripe MCP docs cover both auth paths and the connected-account header.

7. Supabase

Supabase's remote server at https://mcp.supabase.com/mcp manages projects, runs SQL, deploys Edge Functions, and reads logs. Two query parameters do the safety work: read_only=true runs queries as a read-only Postgres role, and project_ref=<id> limits the agent to one project.

claude mcp add --transport http supabase "https://mcp.supabase.com/mcp?read_only=true&project_ref=<id>"

Add write access only after the read-only session has proven the agent asks the questions you'd ask.

8. Cloudflare

Cloudflare publishes a set of remote servers, each on its own hostname. The two that earn a permanent slot are Documentation at https://docs.mcp.cloudflare.com/mcp, which answers Workers and bindings questions against current reference material, and Observability at https://observability.mcp.cloudflare.com/mcp, which searches your Worker logs and analytics.

claude mcp add --transport http cloudflare-docs https://docs.mcp.cloudflare.com/mcp

If you build on Workers, this is the difference between the agent guessing at a Durable Object API and reading it. Our own remote MCP server guide shows the other side of that relationship: what it takes to publish a Workers-hosted MCP server with OAuth.

9. Vercel

Vercel MCP at https://mcp.vercel.com searches Vercel's docs, manages projects and deployments, reads deployment logs, and queries Web Analytics. Its most common use is the failed-deploy loop: the agent pulls the build log for the broken deployment, fixes the cause, and confirms the next deployment went green, without you copying logs out of the dashboard.

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

Sign-in is OAuth through /mcp, and Vercel only accepts MCP clients it has reviewed, with Claude Code on that list. Connecting grants the agent the same access as your Vercel user, so Vercel recommends keeping human confirmation on for writes.

10. Figma

Figma's remote MCP server at https://mcp.figma.com/mcp gives the agent the design context behind a frame: variables, components, and layout data, plus Code Connect mappings so generated code reuses your real components instead of approximating them. For an app team that designs screens in Figma before building them, it closes the gap where the agent guesses at spacing and color tokens it could have read.

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

Sign-in is Figma's OAuth flow through /mcp. Then select a frame in Figma and paste its link into the prompt; the server resolves the node and returns its context. Check Figma's docs for the current seat requirements before rolling it out to a whole team.

11. Exa

Exa's server adds semantic web search built for agents, which returns full page content rather than ten blue links. Claude Code's built-in web search covers casual lookups; Exa is for research tasks where the agent needs to read a dozen sources and compare them. The hosted endpoint is https://mcp.exa.ai/mcp and it needs an API key, passed as a header.

claude mcp add --transport http exa https://mcp.exa.ai/mcp --header "x-api-key: <key>"

12. Executor

Executor is an MCP gateway rather than a single integration: it sits in front of the servers, OpenAPI specs, and GraphQL APIs you connect and exposes them to Claude Code as one endpoint with one tool. That matters once the list above grows, because every connected server adds its full tool list to each request; Executor keeps that to one tool and executes calls in a sandbox so credentials never reach the model. It's MIT licensed, with a local CLI, a desktop app, and a hosted cloud tier that's free for up to three people.

npm install -g executor && executor install
npx add-mcp http://127.0.0.1:4788/mcp --transport http --name executor

The local service serves a streamable HTTP endpoint on port 4788, and the Connect card in its web UI shows the exact command for your setup. Pick the hosted tier if you'd rather not run the service yourself.

13. Screenies

Screenies runs a hosted MCP server at https://screenies.app/mcp for one narrow job: turning your real app screenshots into a finished App Store or Google Play screenshot set from inside your agent. The tools cover uploading captures, generating the set, polling the batch, editing headlines, reordering slides, localizing captions, and checking your credit balance.

claude mcp add --transport http screenies https://screenies.app/mcp

Sign-in is OAuth 2.1 through /mcp, the agent never sees a password, and access is revocable from your account. It generates iPhone, iPad, Android phone, and Android tablet sets at one device and orientation per batch, and the source UI in each slide is your uploaded capture composited into the device frame in code, so the image model can't invent app screens. The server is included with every account and spends the same one-time credits as the dashboard, one per phone slide and two per tablet slide, with headline edits free. Follow the Claude Code setup page for the connect flow, and the Claude Code screenshot walkthrough for an end-to-end example from upload to download.

Servers this list leaves out on purpose

Filesystem and terminal servers duplicate Claude Code's built-in tools and widen the permission surface for no gain. Slack's community server needs a bot token pasted into your config, so it's better handled through Claude's own connectors if you need it at all. "Sequential thinking" servers add a reasoning scratchpad that current models don't need. And the directories that rank for this search list thousands of servers with little curation; treat them as a search index, not a recommendation.

Keep the set small

A dozen connected servers means a dozen tool lists in every request, and the agent gets worse at picking the right one as the count grows. A gateway like Executor is one answer; the other is restraint. Add the servers for the systems you touch weekly, put the shared ones in a committed .mcp.json so the whole team gets the same setup, and use each service's read-only mode until the agent has shown it uses the tools the way you would. When a server stops earning its place, claude mcp remove <name> takes it out.

Questions about MCP servers in Claude Code

What's the difference between --transport http and a stdio server?

HTTP servers are hosted by the vendor, so nothing runs on your machine and OAuth handles sign-in. Stdio servers are local processes Claude Code starts for you, usually with npx, and they take environment variables for keys.

Should I use --scope project or --scope user?

Use project scope for servers the whole team should share on that repository, since it writes a .mcp.json you can commit. Use user scope for servers you want in every project, like Context7 or Playwright.

Do any of these servers cost money?

None charge for MCP access itself. Sentry, Linear, Notion, Stripe, Supabase, Cloudflare, Vercel, and Figma work on their existing plans, Exa needs an API key on its usage pricing, Executor's hosted tier is free for small teams, and Screenies spends the same credits as its dashboard.

How do I revoke an agent's access?

For OAuth servers, revoke from the vendor's account settings (Stripe and Screenies both list connected sessions), then claude mcp remove <name>. For servers that take a key, rotate the key you passed in.

Why isn't the Filesystem server on the list?

Claude Code already reads and edits files inside your project with its built-in tools and asks before going outside it. Adding a filesystem server mostly grants the agent broader disk access than you intended.

Can the same servers be used from Cursor or Claude.ai?

Yes. Every HTTP server above takes the same URL in any client that supports remote MCP with OAuth. The stdio servers work wherever the client can launch a local command.