Best 10 MCP Servers for Claude in 2026: Top Picks

There are thousands of MCP servers you can connect to Claude. Many of them do the same job in slightly different ways, and plenty turn out to be less useful than they look.

Here are 10 that each do a different job, and most are built by the company behind the product. Start with the ones that match how you work.

How We Picked These MCP Servers

Every server on this list meets four criteria:

Actively maintained: recent releases and commits, not a repository that stopped updating a year ago.
Official or vendor-backed where possible: built by the company that owns the product, so it keeps up with that product's API.
Works with Claude: documented setup for Claude Desktop, Claude Code, or both.
Does a distinct job: no two servers on the list solve the same problem.

The Best MCP Servers for Claude at a Glance

No.	MCP server	Best for	Built by	How it connects
1	MemoryPlugin	Memory across sessions and AI tools	MemoryPlugin	Hosted or local
2	Context7	Up-to-date library and API docs	Upstash	Hosted or local
3	Filesystem	Reading and editing local files	MCP project	Local
4	GitHub	Repos, issues, and pull requests	GitHub	Hosted or local
5	Firecrawl	Scraping and crawling websites	Firecrawl	Hosted or local
6	Tavily	Real-time web search	Tavily	Hosted or local
7	Playwright	Browser automation and testing	Microsoft	Local
8	Notion	Your workspace docs and databases	Notion	Hosted
9	Supabase	Databases and backend projects	Supabase	Hosted
10	Linear	Issues, projects, and cycles	Linear	Hosted

1. MemoryPlugin: Memory Across Claude Sessions and AI Tools

Claude's built-in memory only covers Claude. MemoryPlugin is a memory layer that keeps your memory outside your AI tools. It works across 21+ AI tools, including ChatGPT, Claude, and Perplexity, and keeps the same memory in sync between them.

Save something in Claude, and it's there the next time you open ChatGPT or Codex.

Your memories are a plain list of entries you can read, edit, and delete in one place. Through MCP, Claude pulls them into your chats, so your role, your projects, and your preferences come along, and you control exactly which memories it sees.

On top of that, Claude gets:

Memories organized in buckets: keep work, personal, and each project separate, so Claude only gets the bucket you pick. With Smart Memory on, each bucket is also grouped into categories, so only what's relevant loads.
Chat history recall: Claude can search your past conversations, including chats imported from ChatGPT and other tools, and get back a short summary instead of a full transcript.
File search: Claude can query documents you've uploaded to MemoryPlugin.

How to add it to Claude: connect the hosted Remote MCP server at https://www.memoryplugin.com/api/mcp/mcp and approve access in your browser, with nothing to install. The Claude setup guide also covers the local server.

2. Context7: Up-to-Date Docs for Any Library

Claude's knowledge of a library stops at its training data, so it can suggest APIs that changed or never existed. Context7, from Upstash, pulls current, version-specific documentation and code examples into your prompt instead.

It indexes more than 153,000+ libraries, including Next.js, React, Supabase, and Playwright, and its GitHub repository has over 63,000 stars.

It works through two tools:

resolve-library-id: finds the right library from a name like "next.js."
query-docs: fetches the relevant docs and examples for your question.

How to add it to Claude: run npx ctx7 setup --claude, which signs you in and sets everything up. You can also connect manually to https://mcp.context7.com/mcp with your API key. A free key from the Context7 dashboard gives you higher rate limits.

Worth knowing: add "use context7" to a prompt to trigger it, or add a rule to your CLAUDE.md so Claude uses it whenever you ask about a library.

3. Filesystem: Read and Edit Files on Your Computer

The Filesystem MCP server is one of the reference servers from the MCP project itself, with nearly 440,000 weekly downloads on npm. It lets Claude read, write, search, and organize files in folders you choose.

Its tools cover the basics you'd expect:

Reading: read text files, images, or several files at once.
Editing: write files, or make targeted edits with a dry-run preview first.
Organizing: create, list, move, and search files and directories.

How to add it to Claude: add it to your claude_desktop_config.json with npx -y @modelcontextprotocol/server-filesystem, followed by the folders Claude is allowed to access. It needs Node.js 24 or later when run with npx.

Worth knowing: this server is most useful in Claude Desktop. Claude Code already reads and edits files in your project, so you only need Filesystem there for folders outside your working directory. Access is always limited to the directories you allow.

4. GitHub: Repos, Issues, and Pull Requests

The GitHub MCP Server is GitHub's official server, with over 33,000 stars. It gives Claude access to your repositories, issues, pull requests, and GitHub Actions.

You can use it to:

Work with code: browse repositories, search files, and analyze commits.
Manage issues and PRs: create, update, triage, and review them.
Check CI: monitor workflow runs and look into failed builds.

How to add it to Claude: in Claude Code, connect to GitHub's hosted server at https://api.githubcopilot.com/mcp/ with a personal access token. In Claude Desktop, run the local server from claude_desktop_config.json, since the hosted server can't be added as a Claude connector yet. GitHub's Claude installation guide covers both.

Worth knowing: on the hosted server, pick toolsets with URLs like /mcp/x/issues and add /readonly to block changes. The local server uses the --toolsets and --read-only flags. Lockdown mode also filters content from people without push access on public repos, which reduces the risk of prompt injection through issues and comments.

5. Firecrawl: Scrape and Crawl Any Website

Claude can't read most web pages on its own. Firecrawl turns any URL into clean Markdown or structured JSON that Claude can work with.

The full server has 27 tools, including:

Scrape: pull one page as Markdown or as JSON that matches a schema you define.
Map and crawl: find every URL on a site, or extract content from many pages.
Search: search the web and fetch page content in one call.
Agent: research across several sites and return structured data.

How to add it to Claude: install the native Firecrawl connector in one click for search and scraping (8 tools). For the full set, including map, crawl, and agent, connect to https://mcp.firecrawl.dev/v2/mcp-oauth and sign in through your browser. Firecrawl's MCP docs cover every option.

Worth knowing: you can try it with no account. The keyless endpoint at https://mcp.firecrawl.dev/v2/mcp offers scrape, search, and parse within daily limits.

6. Tavily: Real-Time Web Search Built for AI

Where Firecrawl is strongest at pulling content from pages you already have in mind, Tavily is built for search. It returns results formatted for AI, so Claude gets clean, relevant answers to cite instead of raw search pages.

It has five tools:

tavily_search: real-time web search.
tavily_extract: pull data from specific pages.
tavily_map: build a structured map of a website.
tavily_crawl: explore a site systematically.
tavily_research: multi-source research that returns a written answer.

How to add it to Claude: in Claude Code, run claude mcp add tavily --transport http https://mcp.tavily.com/mcp/ and sign in when prompted. In Claude Desktop, click Add to Claude on Tavily's MCP page, or add the URL as a custom connector. Tavily's MCP docs cover other clients.

Worth knowing: you can set default search parameters, like the number of results or search depth, so you don't have to repeat them in every prompt.

7. Playwright: Browser Automation and Testing

Playwright MCP, from Microsoft, lets Claude control a real browser: open pages, click, fill forms, take screenshots, and check network requests. Its GitHub repository has over 38,000 stars.

Instead of reading screenshots, it works from the page's accessibility tree, a structured list of every element on the page. That makes it fast and reliable, and it doesn't need a vision model.

You can use it to:

Test your app: have Claude click through a flow and confirm it works.
Debug the front end: read console messages and network requests.
Reuse logins: keep cookies between sessions, or connect to your existing browser with the Playwright extension.

How to add it to Claude: in Claude Code, run claude mcp add playwright npx @playwright/mcp@latest. For Claude Desktop, add the standard config from the Playwright MCP guide.

Worth knowing: the Playwright team notes that coding agents can save tokens with the Playwright CLI and skills instead. The MCP server is the better fit for long, exploratory sessions where Claude needs to keep the browser open.

8. Notion: Your Workspace Docs and Databases

If your team's knowledge lives in Notion, the Notion MCP server lets Claude search it, read pages, and update them.

Notion recommends its hosted Remote Notion MCP, which offers:

Semantic search: find content across your workspace, and connected apps on plans with Notion AI, by meaning, not only keywords.
Markdown editing: read and edit pages in Markdown, which uses fewer tokens than Notion's raw block format.
Your existing permissions: OAuth sign-in respects what each person can already see in Notion.

How to add it to Claude: follow Notion's MCP setup guide to connect the hosted server. It uses OAuth, so there are no API tokens or JSON files to manage.

Worth knowing: skip the older self-hosted Notion server on GitHub. Notion says it's no longer actively maintained and may be retired.

9. Supabase: Databases and Backend Projects

The Supabase MCP server connects Claude to your Supabase projects, so it can list tables, run SQL, apply migrations, read logs, and deploy Edge Functions.

Its tools are grouped by job:

Database: list tables, run queries, and manage migrations.
Debugging: query logs and get security and performance advice.
Development: generate TypeScript types from your schema.
Edge Functions: list, read, and deploy functions.

How to add it to Claude: in Claude Code, run claude mcp add --scope project --transport http supabase "https://mcp.supabase.com/mcp", then run /mcp and choose Authenticate. You sign in through your browser, with no access token to create.

Worth knowing: Supabase recommends connecting to production only when a task needs it, scoped to one project in read-only mode, and using a development branch to test changes. Add read_only=true to run queries as a read-only user, project_ref to limit access to one project, and features to load only the tool groups you need. More detail is in the GitHub repository.

10. Linear: Issues, Projects, and Cycles

The Linear MCP server lets Claude find, create, and update issues, projects, and comments in Linear, so you can triage bugs or plan work without switching apps.

It's hosted and maintained by Linear, and it offers two access levels:

Read and write: connect to https://mcp.linear.app/mcp.
Read only: connect to https://mcp.linear.app/mcp/readonly, which only exposes read tools.

How to add it to Claude: on Claude Free, Pro, and Max, add the Linear connector from Customize > Connectors. On Team and Enterprise plans, connect it from the connectors page. In Claude Code, run claude mcp add --transport http linear-server https://mcp.linear.app/mcp, then /mcp to sign in.

Worth knowing: the older community-built Linear MCP server is deprecated, and its author recommends Linear's official server instead.

How to Add an MCP Server to Claude

Most servers on this list connect in one of four ways, depending on where you use Claude:

Connectors in Claude: for hosted servers, add the server URL under Customize > Connectors, or install it from the connector directory when one exists, like Firecrawl and Linear. Free plans can add one custom connector, and on Team and Enterprise an owner adds them. (GitHub is the exception: in Claude Desktop it runs locally.)
Claude Desktop config file: for local servers, add an entry to claude_desktop_config.json and restart Claude Desktop.
Claude Code: run claude mcp add with the server's name and URL or command, then use /mcp to sign in if the server needs it.
Desktop Extensions: some local servers ship as a one-click .mcpb file you open in Claude Desktop.

Hosted servers are usually the easiest option. There's nothing to install, and sign-in happens in your browser.

How to Use MCP Servers Safely

An MCP server acts with your permissions, so a careless setup can let Claude change or expose more than you intended. A few habits keep the risk low:

Start read-only: GitHub, Supabase, and Linear all offer read-only modes. Use them until you need write access.
Limit the scope: restrict Filesystem to specific folders, scope Supabase to one project, and load only the GitHub toolsets you use.
Watch for prompt injection: content from issues, tickets, or web pages can contain hidden instructions. Keep manual approval of tool calls on, and review what Claude is about to do.
Keep production separate: test changes on development branches, and connect to production only read-only and scoped to one project.
Check the publisher: install from the vendor's official repo or the Claude directory, and avoid look-alike community packages.

Which MCP Servers Should You Install First?

You don't need all 10. Start with the ones that match your work:

For everyone: MemoryPlugin, so Claude remembers you across chats, plus Tavily or Firecrawl for the web.
For developers: GitHub, Context7, and Playwright, plus Supabase if you use it.
For teams: Notion and Linear, so Claude can work with your docs and your backlog.

Give Claude a Memory That Follows You

Most MCP servers give Claude access to something new. MemoryPlugin gives it continuity. Your preferences, decisions, and past conversations carry over to every new chat, in Claude and in the other AI tools you use.

Start your free 7-day trial and connect MemoryPlugin to Claude in a few minutes.