Best MCP Servers for Claude Code and Claude Desktop

Lets Claude open pages, click, type and read the page through the accessibility tree rather than screenshots, which is faster and cheaper. The default set is about 24 tools; extras for tabs, storage, network, PDF and devtools are opt-in with --caps.

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

Context tip: Microsoft's own README says coding agents may be better served by the Playwright CLI plus a skill, which uses fewer tokens. Try the MCP first; switch if context gets tight.

From Google's Chrome DevTools team. Beyond browsing, it records performance traces, reads network and console, runs Lighthouse and takes heap snapshots. This is the one to reach for when Claude needs to debug a slow page or a layout shift. It has around 65 tools, so start it with --slim if you only need to browse.

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

Fetches current, version-specific documentation for thousands of libraries on demand. Only two tools, so it is almost free in context, and it fixes the most common AI coding failure: code written for last year's version of a framework. A free API key raises the rate limit.

npx ctx7 setup --claude

Or install it as a plugin: /plugin install context7@claude-plugins-official.

One of the seven reference servers still maintained by the MCP project. It fetches a URL and converts it to Markdown. Most useful in Claude Desktop, since Claude Code already has web fetch built in. Its README warns it can reach local and internal IP addresses, so be careful on a work network.

claude mcp add fetch -- uvx mcp-server-fetch

Pick one search server. All four are official and MIT licensed. Exa and Tavily are remote with OAuth, so setup is one line. Brave needs an API key but lets you whitelist tools. Perplexity returns researched answers rather than links.

claude mcp add --transport http exa https://mcp.exa.ai/mcp
claude mcp add --transport http tavily https://mcp.tavily.com/mcp

Note that the old @modelcontextprotocol/server-brave-search package is deprecated. Use Brave's own @brave/brave-search-mcp-server.

One server for Postgres, MySQL, MariaDB, SQL Server, Oracle and SQLite, with just two tools (run SQL, search objects), a read-only mode and row limits. Anthropic uses it as the example in the Claude Code docs. This replaces the old reference Postgres and SQLite servers, which are archived.

claude mcp add db -- npx -y @bytebase/dbhub --dsn "postgresql://user:pass@localhost:5432/app"

Query tables, apply migrations, read logs and search Supabase docs. Remote, with OAuth. Lock it down with URL options: read_only=true, project_ref= to pin one project, and features= to limit tools. Supabase's own docs warn about prompt injection and say not to connect it to production.

claude mcp add --scope project --transport http supabase "https://mcp.supabase.com/mcp?read_only=true"

• MongoDB: npx -y mongodb-mcp-server@latest setup. Use --readOnly; needs Node 22.13 or later.
• Neon: remote at https://mcp.neon.tech/mcp, add ?readonly=true. Neon says not to use it in production. The old npm package is deprecated.
• PlanetScale: claude mcp add --transport http planetscale https://mcp.pscale.dev/mcp/planetscale. Reads go to replicas, and DELETE without WHERE is blocked.

Read and manage repos, issues, pull requests, Actions runs and code scanning alerts. The remote version needs no install:

claude mcp add --transport http github https://api.githubcopilot.com/mcp/ \
  --header "Authorization: Bearer YOUR_GITHUB_PAT"

Context tip: across all toolsets it has 90+ tools. The default set is repos, issues, pull requests, users and context. Add /readonly to the URL, or use a toolset-specific URL, to keep it small. Honestly, for most work the gh CLI (which Claude Code can already run) is lighter. The old @modelcontextprotocol/server-github package is deprecated.

Gives Claude the actual layout, components, variables and styles of a Figma frame, so generated UI matches the design instead of approximating a screenshot. The remote server works on every plan; the desktop server needs a paid Dev or Full seat. Writing back to the canvas is in beta.

claude plugin install figma@claude-plugins-official

Claude can read issues, stack traces and releases, and ask Sentry's Seer for a root cause. The loop "look at the new error, find the bug, fix it" becomes one prompt. Self-hosted Sentry works through the local npm server.

claude mcp add --transport http sentry https://mcp.sentry.dev/mcp

Cloudflare's "Code Mode" server covers about 2,500 API endpoints while using around 1,000 tokens of context, because Claude writes small code snippets against the API instead of loading a tool per endpoint. That makes it one of the most context-efficient large servers available. Product-specific servers (Workers, DNS, logs, browser rendering) also exist.

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

Both are remote with OAuth. Claude reads the ticket or spec, does the work, then updates the issue.

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

Linear offers a /mcp/readonly endpoint. Notion's old local server is no longer maintained; use the remote one.

Stripe (https://mcp.stripe.com) can search docs, create test products and look up payments. Vercel (https://mcp.vercel.com, public beta) can read deployments and logs, and can now also deploy and make purchases. Both are powerful. Use test mode and a restricted account.

Lets Claude Desktop read and write files, limited to the folders you list. This is what turns the Desktop app into a basic local assistant. Claude Code does not need it.

A simple knowledge graph Claude can write facts into and recall later. Useful in Desktop for long-running personal projects. In Claude Code, CLAUDE.md files and auto memory usually do this job better.

Docker's gateway runs each MCP server in its own container, with managed secrets, from a catalogue in Docker Desktop. One command connects the whole set to Claude Code: docker mcp client connect claude-code --global. It is the tidiest way to run many servers without Node and Python installs everywhere.

Home Assistant ships an MCP server built in. Claude can read sensors and control only the devices you have exposed to Assist. More in our MCP servers for home lab guide.