Best MCP Servers for Claude Code (Coding, Docs & Browsers)

Short version: the best MCP servers for Claude Code are the ones that fix a specific gap, not the ones with the longest tool lists. For most developers that means three jobs: up-to-date docs (Context7), your code host (GitHub's official server), and a real browser (Playwright MCP or Chrome DevTools MCP). Add Firecrawl when you need the open web, then Sentry, Supabase, Figma, Linear or Notion only when your work actually lives there.

This roundup is built from Anthropic's Claude Code MCP docs, the Model Context Protocol site and each server's own repository or setup page, all checked on 21 October 2026. Every install command below is copied from those sources. We haven't run benchmark tests on these servers, so there are no speed scores or "we tested" claims here: the picks are our editorial take on what each server is built to do. If you're still choosing your coding agent, read our Claude Code vs Google Antigravity comparison first.

What an MCP Server Does Inside Claude Code

MCP (Model Context Protocol) is an open-source standard for connecting AI applications to external systems. The MCP site describes it as a USB-C port for AI apps: one plug shape, many devices. Claude Code speaks MCP natively, so a server written once works in Claude Code, Cursor, VS Code and every other MCP client.

The architecture is simple. Claude Code is the MCP host. For every server you connect, it creates one MCP client, and each client keeps a dedicated connection to its MCP server. The server is the program that exposes tools (actions like "create issue"), resources (data like a file or a doc page) and prompts. Behind the server sits the actual tool or data: your GitHub repos, a Chrome window, a Postgres database, a docs index.

Anthropic's docs put the use case plainly: connect a server when you find yourself copying data into chat from another tool. Instead of pasting an error from Sentry or a spec from Notion, Claude reads it and acts on it directly. That's also why MCP is a context engineering decision, not just a plugin choice: every server changes what Claude can see.

Servers run in two ways. Local stdio servers run as a process on your machine (usually launched with npx, uvx or Docker) and typically serve one client. Remote HTTP servers are hosted by the vendor, usually sign in with OAuth, and serve many clients. Most vendor servers in this list are now remote.

How to Add an MCP Server to Claude Code

Everything goes through the claude mcp command, run in your terminal (not inside a Claude session). The docs call HTTP "the recommended option for connecting to remote MCP servers":

# Remote HTTP server
claude mcp add --transport http <name> <url>

# Remote server with a token header
claude mcp add --transport http secure-api https://api.example.com/mcp \
  --header "Authorization: Bearer your-token"

# Local stdio server: everything after -- is the launch command
claude mcp add [options] <name> -- <command> [args...]

Three transports matter. HTTP (streamable HTTP) is the default for hosted servers. SSE is deprecated; recent Claude Code versions try HTTP first and fall back to SSE on their own, and --transport sse still exists for older setups. stdio is for local processes. The -- separator matters: it stops Claude Code from reading the server's own flags as its options.

Scopes: local, project and user

Add --scope project to write a shared .mcp.json your team commits, or --scope user for servers you want everywhere (Context7 is a natural user-scope server). When the same name exists in more than one scope, local wins, then project, then user.

Checking and signing in

Use claude mcp list to see every server with a health status, claude mcp get <name> for details and claude mcp remove <name> to delete one. Inside a session, /mcp shows status, tool counts and the OAuth sign-in for remote servers; you can also run claude mcp login <name> from your shell. Claude Code stores the tokens and refreshes them, and "Clear authentication" in /mcp revokes access.

Pro tip: Don't worry too much about tool-list bloat. Claude Code's tool search defers MCP tool definitions, so only tool names and server instructions load at session start. Large tool outputs still cost context: Claude Code warns when one MCP result passes 10,000 tokens and caps results at 25,000 tokens by default (MAX_MCP_OUTPUT_TOKENS changes it). Pair that with prompt caching if long agent sessions are getting expensive.

Quick Comparison: The Best MCP Servers for Claude Code

What we looked for

• An official or clearly maintained source. Anthropic says it reviews connectors for its Directory but "does not security-audit or manage any MCP server." Prefer the vendor's own server.
• Remote HTTP with OAuth where possible. No local process to babysit, no long-lived token in a config file.
• Scoping controls. Project scoping, read-only modes and feature groups (Supabase and Sentry both have them) shrink what a bad prompt can do.
• A gap Claude Code doesn't already fill. Claude Code already reads files, edits code and runs shell commands, so a server should add something new: live docs, a browser, your issue tracker.
• Sensible output size. Servers that return clean Markdown or structured data keep your context window usable.

The servers side by side

Our Top Pick: Context7 for Up-to-Date Docs

Honest answer: the most common way a coding agent wastes your time is writing code against an API that changed after its training data. Context7, from Upstash, targets exactly that. It pulls up-to-date, version-specific documentation and code examples from the source and puts them into Claude's context, so you get fewer invented methods and fewer deprecated patterns.

You use it by adding "use context7" to a prompt, or by naming a library ID directly (for example /supabase/supabase) so it skips the matching step. Mention a version ("Next.js 16") and it matches that version's docs. Context7's Claude Code setup page gives both forms, and it recommends a free API key for higher rate limits:

# Remote server (user scope, all projects)
claude mcp add --scope user --header "Authorization: Bearer YOUR_API_KEY" --transport http context7 https://mcp.context7.com/mcp

# Local server
claude mcp add --scope user context7 -- npx -y @upstash/context7-mcp --api-key YOUR_API_KEY

There's also npx ctx7 setup --claude, which signs you in with OAuth, creates the key and lets you choose between MCP mode and a CLI-plus-skill mode. My take: install it at user scope and add a line to your CLAUDE.md telling Claude to use Context7 for any library question. It's the one server that improves almost every coding session.

GitHub MCP Server: Best for PRs and Issues

GitHub's official MCP server gives Claude Code access to repositories, pull requests and issues, so prompts like "Review PR #456 and suggest improvements" or "Show me all open PRs assigned to me" (both examples from Anthropic's docs) work without copying anything into chat. The older GitHub server in the MCP reference repo is archived; this is the one to use.

The remote server lives at GitHub's Copilot API endpoint and authenticates with a fine-grained personal access token. GitHub's Claude install guide gives this command for Claude Code 2.1.1 and newer:

claude mcp add-json github '{"type":"http","url":"https://api.githubcopilot.com/mcp","headers":{"Authorization":"Bearer YOUR_GITHUB_PAT"}}'

For older versions, the same guide gives claude mcp add --transport http github https://api.githubcopilot.com/mcp -H "Authorization: Bearer YOUR_GITHUB_PAT". If you'd rather run it locally, there's a Docker image (ghcr.io/github/github-mcp-server) with a browser OAuth login or a PAT. Two habits from GitHub's own guide are worth copying: keep the token in a .env file, and add both .env and .mcp.json to .gitignore if the config holds a token.

My take: give the token access only to the repositories Claude should touch. A fine-grained PAT scoped to one repo is the cheapest safety upgrade on this list.

Playwright MCP: Best for Browser Automation

Playwright MCP from Microsoft lets Claude drive a real browser through Playwright. The key design choice is that it works from the page's accessibility tree, not screenshots, so Claude acts on structured data and doesn't need vision to click the right button. That makes it a good fit for "open the signup page, fill the form, tell me what breaks" checks after Claude changes your UI.

claude mcp add playwright npx @playwright/mcp@latest

It needs Node.js 18 or newer and opens a headed browser by default (--headless changes that). --isolated keeps the browser profile in memory instead of on disk, which is what you want for throwaway test sessions. Note the README's own warning: its --allowed-origins and --blocked-origins flags are "not a security boundary".

One honest caveat from Microsoft itself: the README says coding agents may get better token efficiency from the separate Playwright CLI with skills, because MCP loads accessibility trees into context. MCP still suits longer exploratory loops where keeping browser state matters. If you need Claude to operate a whole desktop rather than a browser, that's a different tool: see our guide to Claude computer use.

Chrome DevTools MCP: Best for Debugging Live Pages

Where Playwright is about doing things in a browser, Chrome DevTools MCP is about understanding what the browser is doing. It lets Claude control and inspect a live Chrome instance: record performance traces and pull out insights, read network requests, take screenshots and check console messages with source-mapped stack traces. Automation runs through Puppeteer.

claude mcp add chrome-devtools --scope user npx chrome-devtools-mcp@latest

You can also install it as a Claude Code plugin with skills: run /plugin marketplace add ChromeDevTools/chrome-devtools-mcp, then /plugin install chrome-devtools-mcp@chrome-devtools-plugins. Read the disclaimers before you use it on a logged-in profile: the server exposes browser content to the MCP client, Google collects usage statistics by default (--no-usage-statistics opts out), and performance tools may send trace URLs to the CrUX API (--no-performance-crux turns that off). It officially supports Google Chrome and Chrome for Testing.

My take: install Playwright for "click through this flow" and Chrome DevTools for "why is this page slow, and what's throwing in the console". Many frontend developers will want both.

Firecrawl MCP: Best for Web Search and Scraping

When the answer isn't in your repo or your docs, Firecrawl brings in the open web. Its MCP server can search the web, scrape any URL into clean Markdown or structured JSON, parse PDFs, crawl or map a whole site, and interact with pages in a real browser (click, type, fill forms). It also has search indexes for developer sources such as GitHub issues, merged pull requests, READMEs and docs, which is handy when an error message only has answers in a closed issue.

Firecrawl's docs give two Claude Code commands. The OAuth endpoint unlocks the full tool set after you sign in via /mcp; the keyless endpoint works with no account but only exposes scrape, search and parse, rate-limited:

# Sign in with OAuth (full tool set)
claude mcp add --transport http firecrawl https://mcp.firecrawl.dev/v2/mcp-oauth

# Keyless free tier (scrape, search, parse only)
claude mcp add --transport http firecrawl https://mcp.firecrawl.dev/v2/mcp

Firecrawl's own rule is worth repeating: never put an API key in the server URL or in a chat; configure it as a header or secret. Crawls use credits, so check our Firecrawl pricing breakdown before you let an agent crawl a large site. Because Firecrawl fetches external pages, it's also exactly the kind of server Anthropic flags for prompt injection risk, so keep tool approvals on when you scrape unknown sites.

Sentry MCP: Best for Fixing Production Errors

Sentry's MCP server is a remote service designed, in Sentry's words, for human-in-the-loop coding agents. It lets Claude search errors, analyse performance and triage issues, which turns "check Sentry for the crash in checkout and fix it" into one prompt. All connections use OAuth.

claude mcp add --transport http sentry https://mcp.sentry.dev/mcp/{organizationSlug}/{projectSlug}

Sentry recommends scoping the URL to one organisation and project when you can: tools then default to that project and unnecessary discovery tools are hidden. Anthropic's own docs use the unscoped https://mcp.sentry.dev/mcp as their sign-in example, which works too. There's also a Claude Code plugin (claude plugin marketplace add getsentry/sentry-mcp, then claude plugin install sentry-mcp@sentry-mcp) that adds a Sentry subagent Claude delegates to.

Supabase MCP: Best for Database Work

The Supabase MCP server connects Claude to your Supabase projects to manage tables, fetch config and query data. Supabase's Claude Code command adds it at project scope:

claude mcp add --scope project --transport http supabase "https://mcp.supabase.com/mcp"

This is the server where security settings matter most, and Supabase says so first. Append ?project_ref=<id> to limit it to one project, ?read_only=true to run queries as a read-only Postgres user, and ?features=database,docs to enable only the tool groups you need. Supabase's docs walk through a prompt injection example (a support ticket containing hidden SQL instructions) and recommend keeping manual tool approval on, avoiding production data unless the task needs it, using branching for safe changes, and never giving the server to end users.

Figma MCP Server: Best for Design to Code

The Figma MCP server gives agents the design context they need to generate code from Figma files, and the remote server can also write native Figma content back to the canvas. Figma lists Claude Code as supporting both its desktop and remote servers, plus a Figma plugin. To add the remote server manually, Figma's developer docs give:

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

Add --scope user to make it available in every project. Figma's guide lists it for Starter, Professional, Education, Organization, Government and Enterprise plans. Use it when you build from real design files; skip it if your UI work starts from a text spec.

Linear and Notion: Best for Specs and Issues

If your tickets live in Linear or your specs live in Notion, these two remote servers remove the copy-paste step. Both are added with one HTTP command and then authenticated through /mcp, according to their official setup pages:

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

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

Anthropic uses Notion as its own example of a remote HTTP server, so it's a good first server to practise the OAuth flow on. The real win is the workflow Anthropic describes: "Add the feature described in [ticket] and create a PR on GitHub" only works when both the tracker and GitHub are connected.

Reference Servers: Filesystem, Git and Fetch

The modelcontextprotocol/servers repository holds a small set of reference servers maintained by the MCP steering group: Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking and Time. The repo is blunt that they are "reference implementations" and educational examples, "not as production-ready solutions". It points people looking for real servers to the official MCP Registry.

• Filesystem (@modelcontextprotocol/server-filesystem) only touches the directories you pass on the command line. Claude Code already has native file tools inside your project, so this is mostly useful for reaching a folder outside it.
• Git (mcp-server-git, run with uvx) reads, searches and manipulates repositories. The README says it's in early development, and Claude Code can already run git in the shell.
• Fetch (mcp-server-fetch, run with uvx) turns web pages into model-friendly text. Its README carries a caution that it can reach local or internal addresses if you turn off its private-IP guard.

My take: install these to learn how MCP works, not as your daily stack. For search specifically, the archived reference Brave server has been replaced by Brave's official Brave Search MCP server, which runs locally with a BRAVE_API_KEY; its README gives an npx -y @brave/brave-search-mcp-server --transport stdio launch command that you can wrap with claude mcp add using the -- pattern above.

The Starter Stack: Coding, Docs and Browsers

If you only add a handful of servers this week, group them by the three jobs Claude Code can't do well alone. Our pick for a starter stack, with the scope we'd use for each:

That's three servers to start, six to eight at most for a busy full-stack setup. If you also generate images from Claude, our roundup of the best MCPs for generating images with Claude covers that side of the stack.

Which MCP Server Should You Add First?

MCP Security: Scopes, Tokens and Prompt Injection

Every server you add is new reach for Claude, and the risk grows with both what the server can touch and what untrusted text it reads. Anthropic's docs are direct: verify you trust each server before connecting it, because servers that fetch external content can expose you to prompt injection. Here is what the docs give you to manage that:

• Project servers need approval. In interactive sessions, Claude Code asks before using servers from a project's .mcp.json. A cloned repository can't approve its own servers: approval settings committed to the repo are ignored until you trust the folder. claude mcp reset-project-choices resets your answers.
• Headless runs skip the prompt. In claude -p runs and Agent SDK sessions, project-scoped servers load without asking. For unattended jobs (like the scheduled runs in our Claude Routines explainer), use disabledMcpjsonServers or --strict-mcp-config to control exactly what loads.
• Keep tokens out of git. .mcp.json supports ${VAR} and ${VAR:-default} expansion in commands, args, env, URLs and headers, so the shared file holds variable names, not secrets. Claude Code also blanks its own credentials (such as ANTHROPIC_API_KEY) if a remote server's URL or headers reference them.
• Use the narrowest scope the vendor offers. Fine-grained GitHub tokens, Sentry project URLs, Supabase read_only and project_ref.
• Keep manual approval for tools that write. Supabase's guidance applies to every server: review each tool call, and treat data returned by a tool as information, never as instructions.

Common Mistakes When Adding MCP Servers

• Installing ten servers on day one. More servers means more tool choices and more places for injected text to come from. Start with the three-server stack and add one at a time.
• Forgetting the -- on stdio servers. Without it, Claude Code tries to parse the server's flags as its own options.
• A url with no type in JSON. Claude Code reads an entry with no type as a stdio server, so a pasted config from another client fails until you add "type": "http".
• Committing a token in .mcp.json. Use environment variable expansion, or keep token-bearing servers at local scope.
• Using reference servers as production tools. The MCP repo itself says they aren't production-ready.
• Connecting a production database with write access. Start Supabase in read-only mode, scoped to one project.
• Ignoring output size. A full-site crawl or a huge accessibility tree can fill your context and your bill. Check our Claude Code API pricing guide if long MCP-heavy sessions are running up token costs.

The Verdict

The best MCP servers for Claude Code are the ones that close a gap, not the ones with the longest list of tools. Our pick for almost everyone is Context7 for docs, the official GitHub MCP Server for code, and Playwright MCP for the browser, with Chrome DevTools MCP added for frontend debugging. Reach for Firecrawl when answers live on the open web, and add Sentry, Supabase, Figma, Linear or Notion only when your work already lives there.

Whatever you install, prefer official remote servers with OAuth, pick the narrowest scope, keep tokens out of git and leave tool approvals on for anything that writes data or reads untrusted pages. A small stack you understand beats a big one you don't.

Frequently Asked Questions

Keep Reading

• How to use Claude Code for free: the complete guide
• 25 best Claude Opus 5.5 prompts for coding
• AI skills vs prompts: what's the difference?