Best MCP servers in 2026: the 20 most popular, by category

Top MCP servers

The 20 most popular MCP servers in 2026 for development, design, marketing, product teams and web data, what each one does, and which ones you no longer need

An MCP server gives an AI agent new tools: open a browser, read a Figma file, search your Jira, query Search Console, take a payment. There are more than 10,000 of them now, and most aren't worth installing. This series covers the ones people actually use, tested in Claude Code where we could.

First, the distinction that decides what you need:

Operational MCPs are for your team. You add them to Claude Code, Cursor or Claude Desktop, and the agent does your work with them: tests the app, files the ticket, writes the SEO report.

Integration MCPs are for your product. Your backend attaches them to an API call, and your product's AI feature gets web search, scraping or payments without you writing each integration.

Pick your category

Every server here has its own page in the MCP directory, with setup, prompts and when to pick something else.

The 20 most popular

No single number measures MCP popularity, so we combined three: GitHub stars, weekly visitors on PulseMCP (a public directory of MCP servers that estimates traffic for each one) and monthly Google searches for "name MCP". Hosted servers have no stars, so for those, traffic and searches count more.

#	MCP	What your agent gets	Category	Signal
1	Playwright	A real browser to click through your app	Development	38K ★, 82K searches
2	Context7	Current docs for any library	Development	63K ★
3	Figma	Designs as data: layers, variables, components	Design	74K searches
4	GitHub	PRs, issues and CI logs	Development	34K ★, 69K searches
5	Chrome DevTools	Console, network, performance traces	Development	53K ★
6	Atlassian	Jira issues and Confluence pages	Product	40K searches
7	Storybook	Your component library's docs and tests	Design	2.5M visits/week
8	Browser Use	A cloud browser that logs in and fills forms	Development	118K ★
9	Supabase	Your database, migrations and RLS checks	Development	26K searches
10	Notion	Pages and databases	Product	23K searches
11	Serena	Code navigation by symbol, like an IDE	Development	30K ★
12	Slack	Threads, channels and messages	Product	17.7K searches
13	Zapier	9,000+ apps through one server	Marketing	229K visits/week
14	Linear	Issues, projects and cycles	Product	156K visits/week
15	Firecrawl	Any website as clean markdown	Web data	136K visits/week
16	Stripe	Customers, invoices and refunds	Product	135K visits/week
17	HubSpot	CRM, campaigns and attribution	Marketing	76K visits/week
18	PostHog	Funnels, retention and experiments	Marketing	65K visits/week
19	Ahrefs	Keywords, backlinks and competitors	Marketing	50K visits/week
20	Exa	Web search built for agents	Web data	9K calls on Smithery, an MCP marketplace

How to add one

Where	How
Claude Code	`claude mcp add` for local servers, `claude mcp add --transport http` for hosted ones. Then `/mcp` to sign in.
Claude Desktop and claude.ai	Settings → Connectors. Popular ones are one click; others take a URL. They also show up in Claude Code.
Cursor	Settings → MCP, or `.cursor/mcp.json` in the project.
Codex	`codex mcp add --` , or `--url` for hosted ones.
Your product	Pass the server URL in your API call. See the web data article for code.

Popular, but your agent already does this

Some of the most downloaded servers date from early 2025, before coding agents could do these things themselves. In Claude Code, Codex or Cursor, skip them:

Server	Why you can skip it
Filesystem	Reading and writing files is built into every coding agent
Desktop Commander	So is running terminal commands
Sequential Thinking	Current models reason step by step before answering
Fetch	Web fetch is built in
Memory	Use `CLAUDE.md` or `AGENTS.md` instead

They still help in chat apps that have no file or terminal access.

Three rules before you install

MCP or CLI?

If a tool has a good CLI, like gh, supabase, vercel or stripe, coding agents often do better with the CLI: it's already signed in and adds nothing to context. Use the MCP when the agent needs to see something a CLI can't show (a page, a design, a trace), when the tool has no CLI, or when teammates use chat apps instead of a terminal.