The Best MCP Servers in 2026, by Job (and How to Install Them)

The best MCP servers are the ones published by the vendor whose system they connect to, plus a handful of official reference servers. For most people that means GitHub for code, Playwright or Chrome DevTools for the browser, Context7 or Microsoft Learn for current docs, Notion or Atlassian for work tracking, Sentry for errors, Supabase or Google's MCP Toolbox for databases, and the reference Filesystem, Git and Fetch servers for local work. Every server below was checked against its official repository or docs in October 2026.

Quick picks by job

The official reference servers

The modelcontextprotocol/servers repository now keeps just seven reference servers maintained by the MCP steering group. Its README says they are educational examples, "not as production-ready solutions", so treat them as solid starting points for personal use.

• Filesystem: read, write and search files, limited to directories you pass as arguments. Give it one project folder, never your home directory.
• Git: read, search and manipulate local Git repositories.
• Fetch: retrieve a web page and convert it for the model. Its README warns it "can access local/internal IP addresses", so do not run it on a machine that can reach admin panels you care about.
• Memory: a persistent knowledge-graph memory across sessions.
• Sequential Thinking: a scaffold for step-by-step problem solving.
• Time: time and timezone conversions.
• Everything: a test server exercising every protocol feature, for client developers.

Thirteen older servers were archived, including GitHub, GitLab, Google Drive, PostgreSQL, Puppeteer, Redis, Sentry, Slack and SQLite. Many tutorials still point at @modelcontextprotocol/server-github or server-postgres. Use the vendor's own server instead: GitHub, Sentry and Brave now ship official replacements, and Slack's is maintained by Zencoder. For a published catalog, the repo now points to the MCP Registry.

Best MCP servers for coding

GitHub MCP Server

github/github-mcp-server is GitHub's official server for repositories, issues, pull requests, Actions and more. Its default toolsets are context, repos, issues, pull_requests and users; others, such as actions, are opt-in, which is a good way to keep the tool list short. The remote endpoint is https://api.githubcopilot.com/mcp/, and a Docker image (ghcr.io/github/github-mcp-server) is available for local use with a personal access token. Give the token only the repositories and scopes the task needs.

Context7

Context7 by Upstash feeds current library documentation to coding agents, which cuts down on answers based on outdated APIs. Use the remote endpoint https://mcp.context7.com/mcp, the npm package @upstash/context7-mcp, or run npx ctx7 setup.

Microsoft Learn MCP

MicrosoftDocs/mcp exposes Microsoft's official documentation at https://learn.microsoft.com/api/mcp. Useful if you work with Azure, .NET or Microsoft 365. Microsoft also keeps a catalog of its other official servers at microsoft/mcp.

Best MCP servers for the browser

Playwright MCP

microsoft/playwright-mcp lets the model drive a real browser through Playwright: navigate, click, fill forms, read the page. It is the standard choice for end-to-end testing and form-filling tasks. Run it with npx @playwright/mcp@latest.

Chrome DevTools MCP

ChromeDevTools/chrome-devtools-mcp, described as "Chrome DevTools for coding agents", gives the agent console logs, network requests and performance traces. It is the better pick when you are debugging your own front end. Run npx -y chrome-devtools-mcp@latest; the README shows --headless and --no-usage-statistics options.

Both servers can visit any page, and pages can contain prompt injection. Use a separate browser profile without your logged-in sessions.

Best MCP servers for search and the web

• Brave Search MCP: official replacement for the archived reference server. Package @brave/brave-search-mcp-server, needs a BRAVE_API_KEY.
• Firecrawl MCP: official server for scraping, crawling and search. Needs a Firecrawl API key.
• Exa MCP: web search and crawling from Exa. Needs an Exa API key.

Best MCP servers for work tools

Notion

Notion now recommends its hosted Remote Notion MCP (https://mcp.notion.com/mcp) over its open-source local server, whose README says it "is no longer actively maintained or supported". The remote server uses OAuth and respects each user's existing permissions.

Atlassian (Jira, Confluence)

The Atlassian MCP Server is a remote server for Jira, Confluence and Jira Service Management. Atlassian recommends the v2 endpoint https://mcp.atlassian.com/v2/mcp and says its legacy SSE endpoint stopped being supported after 30 June 2026.

Sentry

Sentry MCP is hosted at https://mcp.sentry.dev/mcp and lets an agent read issues, stack traces and performance data. The MCP docs use it as their example of a remote server.

Best MCP servers for data and cloud

• Supabase MCP: hosted at https://mcp.supabase.com/mcp. It supports a read-only mode; use it unless you truly want the model writing to your database.
• MCP Toolbox for Databases: Google's open-source server for PostgreSQL, AlloyDB, BigQuery and other databases. You define the allowed queries in a config file, which is safer than handing a model raw SQL.
• Stripe: remote server at https://mcp.stripe.com. Use a restricted key.
• Cloudflare: a set of remote servers on *.mcp.cloudflare.com, for example browser, bindings and builds.
• AWS: AWS Labs' collection of open-source servers for AWS services.

How to install an MCP server

The two shapes are the same everywhere: a remote server is a URL (usually with OAuth), and a local server is a command the client launches.

Claude Code

From the Claude Code MCP docs:

# Remote server (recommended)
claude mcp add --transport http notion https://mcp.notion.com/mcp

# Local stdio server: everything after -- is the server command
claude mcp add --transport stdio playwright -- npx @playwright/mcp@latest

# Check what is connected
claude mcp list

Add --scope project to save the server in a shared .mcp.json for your team, or --scope user for all your projects. Inside a session, /mcp shows status and handles OAuth sign-in. Claude Code loads MCP tools on demand through tool search by default, which keeps context small when many servers are connected.

Claude Desktop and claude.ai

For a remote server, go to Customize > Connectors > Add custom connector, paste the server URL and sign in. Free plans can add one custom connector; on Team and Enterprise, an Owner adds it in organization settings. See Anthropic's custom connector guide.

For a local server, install an MCPB desktop extension under Settings > Extensions, or add it to claude_desktop_config.json:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/you/projects/demo"]
    }
  }
}

On Windows, the servers README says to wrap npx with cmd /c.

Cursor

Cursor reads .cursor/mcp.json in a project or ~/.cursor/mcp.json globally, with servers under mcpServers. Remote servers use url; local ones use command, args and env. Values can reference environment variables with ${env:NAME}, so keep keys out of the file.

{
  "mcpServers": {
    "context7": { "url": "https://mcp.context7.com/mcp" },
    "playwright": { "command": "npx", "args": ["@playwright/mcp@latest"] }
  }
}

By default Cursor asks before each MCP tool call (Cursor docs).

VS Code

Run MCP: Add Server from the Command Palette. VS Code now prefers the portable .mcp.json at the workspace root (servers under mcpServers). The older .vscode/mcp.json still works and uses a servers key with a type:

{
  "servers": {
    "github": { "type": "http", "url": "https://api.githubcopilot.com/mcp" }
  }
}

VS Code warns that "Local MCP servers can run arbitrary code on your machine" and shows a trust dialog for servers from outside a trusted workspace (VS Code docs).

MCP server security checklist

The MCP security best practices list real attacks, including malicious startup commands that exfiltrate SSH keys. Before you add a server:

1. Check the publisher. Prefer the vendor's own server or a reference server. Look at the repo, not just a directory listing.
2. Read the exact command. A local server runs with your permissions. If the command contains curl, sudo or anything you do not understand, stop.
3. Prefer remote with OAuth for SaaS tools, so no long-lived key sits in a config file.
4. Scope tokens tightly: read-only database roles, fine-grained GitHub tokens, restricted Stripe keys.
5. Separate reading untrusted content from sending data. A server that browses the web plus a server that can email or post is an exfiltration path if a page injects instructions.
6. Keep approvals on for write tools. Only "always allow" read tools from servers you trust.
7. Turn off what you are not using. Fewer tools means fewer mistakes and a smaller context window bill.

More on protecting agents in AI agent security. New to the protocol? Start with what is MCP, or see how MCP fits with AI agent skills.