Best MCP Servers in 2026: 25 You Should Install Now

Most "best MCP servers" lists dump 50 entries alphabetically. This one is organized by what you actually need to get done -- and every entry was tested in Claude Desktop, Claude Code, or Cursor.

MCP (Model Context Protocol) lets your AI assistant call external tools natively. Instead of copy-pasting data into Claude or Cursor, the assistant pulls it directly -- reading files, querying databases, searching the web, managing deployments. The protocol hit 97 million monthly SDK downloads in March 2026, up from 2 million at launch. Over 20,000 servers sit on registries like Glama and mcp.so, but most are duplicates or weekend experiments. Smithery lists roughly 6,000 distinct servers. Here are the 25 worth installing.

One warning first: 30+ CVEs were filed against MCP servers in January-February 2026 alone. A scan of popular servers found security findings in 66% of them. Every server you install is a potential attack surface. The MCP security guide covers mitigations -- read it before installing anything with write access.

Development & Code

1. GitHub MCP Server

What it does: Full repository management -- commits, branches, PRs, issues, code search -- without leaving your AI assistant.

The most-installed MCP server by a wide margin. If you code with Claude or Cursor, install this first. The official GitHub-maintained server is a Go binary distributed as a Docker image, not an npm package. It supports both read and write operations across repos, issues, pull requests, and code search.

{
  "mcpServers": {
    "github": {
      "command": "docker",
      "args": ["run", "-i", "--rm", "-e", "GITHUB_PERSONAL_ACCESS_TOKEN", "ghcr.io/github/github-mcp-server"],
      "env": {
        "GITHUB_PERSONAL_ACCESS_TOKEN": "ghp_your_token"
      }
    }
  }
}

2. Filesystem MCP Server

What it does: Secure local file operations -- read, write, edit, search, and manage directories with sandboxed access.

The foundational server. Required for any workflow where your AI needs to touch local files. Anthropic-maintained, with directory-level sandboxing -- you specify exactly which paths the server can access. No rogue file reads outside your project.

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

3. Context7 MCP

What it does: Fetches version-specific documentation for any library or framework at call time.

Your AI quotes the current API, not a 2-year-old version from its training data. Context7 resolves library docs on demand. Essential for codebases using fast-moving frameworks (Next.js, Svelte, Tailwind). Built by Upstash, distributed as @upstash/context7-mcp.

{
  "mcpServers": {
    "context7": {
      "command": "npx",
      "args": ["-y", "@upstash/context7-mcp"]
    }
  }
}

4. Playwright MCP

What it does: Browser automation -- navigate pages, click elements, fill forms, take screenshots, run E2E tests.

Turns your AI into a QA engineer. Microsoft-maintained (@playwright/mcp), with full Chromium, Firefox, and WebKit support. More capable than Puppeteer MCP for testing, scraping, and debugging frontend issues. Runs headless by default; pass --headed for visual debugging.

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": ["-y", "@playwright/mcp"]
    }
  }
}

5. Sentry MCP

What it does: Pull error context -- stack traces, breadcrumbs, affected users -- directly into your AI conversation.

Instead of pasting Sentry links, Claude reads the full error context and suggests fixes with the actual stack trace in hand. Published as @sentry/mcp-server. Requires a Sentry auth token with project read access.

{
  "mcpServers": {
    "sentry": {
      "command": "npx",
      "args": ["-y", "@sentry/mcp-server"],
      "env": {
        "SENTRY_AUTH_TOKEN": "sntrys_your_token"
      }
    }
  }
}

6. Docker MCP Server

What it does: Build, run, inspect, and debug containers from natural language.

Ask "spin up a Postgres container on port 5433" and it happens. Also generates Dockerfiles, inspects logs, manages volumes, and lists running containers. Community-maintained as mcp-server-docker. Requires the Docker socket to be accessible.

{
  "mcpServers": {
    "docker": {
      "command": "npx",
      "args": ["-y", "mcp-server-docker"]
    }
  }
}

Web Data & Search

7. Brave Search MCP

What it does: Real-time web search using Brave's independent index.

No knowledge cutoff. Your AI searches the live web and returns current results. The best general-purpose search MCP -- independent from Google, fast, and privacy-respecting. Maintained by Brave (not Anthropic), published as @brave/brave-search-mcp-server. Free tier: 2,000 queries/month.

{
  "mcpServers": {
    "brave-search": {
      "command": "npx",
      "args": ["-y", "@brave/brave-search-mcp-server"],
      "env": {
        "BRAVE_API_KEY": "BSA_your_key"
      }
    }
  }
}

8. Firecrawl MCP

What it does: Scrape any website into clean, LLM-ready markdown.

The gold standard for web scraping via MCP. Converts pages to structured markdown, handles JavaScript rendering, and supports batch crawling of entire sites. Used heavily by Claude and Cursor users building RAG pipelines or doing competitive research.

9. Exa MCP Server

What it does: Semantic search -- find content by meaning, not keywords.

Different from Brave: Exa understands what you're looking for conceptually. "Find recent articles about AI agents that mention distribution" returns relevant results even without exact keyword matches. Particularly strong for research workflows where you know the concept but not the exact phrasing.

Databases & Backend

10. PostgreSQL MCP Server

What it does: Query your Postgres database in natural language.

"Show me the 10 users who signed up this week with the highest engagement" becomes an actual SQL query, executed, with results returned. Read-only mode available for safety. Published as @modelcontextprotocol/server-postgres. Security note: this server has a known SQL injection vulnerability (CVE filed in early 2026) -- always run it in read-only mode against production databases, and consider using Supabase MCP instead for managed Postgres.

11. Supabase MCP

What it does: Manage your entire Supabase project -- database, auth, storage, edge functions -- from your AI.

If your stack includes Supabase, this replaces tab-switching to the dashboard entirely. Create tables, manage RLS policies, debug auth issues, query data. The official Supabase team maintains it, which means it handles connection pooling and auth correctly out of the box.

12. Redis MCP Server

What it does: Inspect and manage your Redis cache through natural language.

Check keys, flush caches, monitor performance, inspect data structures -- all without memorizing Redis CLI commands. Useful for debugging caching issues during development.

Design & Frontend

13. Figma MCP (Dev Mode)

What it does: Exposes the live structure of your Figma selection -- hierarchy, auto-layout, variants, spacing tokens, component refs.

Your AI generates code against the real design, not a screenshot. The gap between "design" and "implementation" shrinks dramatically. Requires Figma Dev Mode access. Particularly effective with Cursor for turning designs into production components in one pass.

14. TailwindCSS MCP

What it does: Provides your AI with your actual Tailwind config -- theme, plugins, custom utilities.

No more hallucinated class names. The AI knows your exact color palette, breakpoints, and custom utilities. Reads your tailwind.config.ts directly.

Deployment & DevOps

15. Vercel MCP

What it does: Monitor deployments, check build logs, manage environment variables, configure domains.

"Why did my last deploy fail?" gets a real answer with the actual build log, not a guess. Also handles environment variable management and domain configuration -- tasks that otherwise require multiple dashboard clicks.

16. Cloudflare MCP

What it does: Manage Workers, KV stores, R2 buckets, and DNS from your AI.

Deploy edge functions, debug Worker errors, manage cache rules -- without touching the Cloudflare dashboard. Maintained by Cloudflare.

Productivity & Integrations

17. Linear MCP

What it does: Read, create, and update issues. Manage sprints. Search across projects.

"Create a bug ticket for the login redirect issue, assign it to me, add it to the current sprint" -- done in one message. Supports full CRUD on issues, projects, cycles, and labels.

18. Slack MCP

What it does: Search channels, summarize threads, draft replies, trigger notifications.

Useful for catching up on conversations you missed, or drafting announcements that match your team's tone. Read-only mode available if you don't want your AI posting on your behalf.

19. Google Drive MCP

What it does: Search, read, and summarize documents across personal and shared drives.

"Summarize the Q1 planning doc" pulls the actual document content, not a hallucinated version. Supports Docs, Sheets, Slides, and PDFs.

20. Composio MCP

What it does: Integration hub -- 250+ platforms (GitHub, Slack, Notion, Jira, HubSpot, etc.) through one MCP server.

Instead of installing 10 separate MCP servers, Composio bundles common integrations through a single composio-mcp package. The trade-off: less depth per integration, more breadth. Good for teams that need light access to many platforms without managing a dozen configs.

Data & Research

21. Toolradar MCP

What it does: Search, compare, and get pricing for 8,400+ software tools with verified data.

The one MCP server that solves the "which tool should I use?" problem. Your AI gets real editorial scores, weekly-verified pricing, G2/Capterra ratings, and AI-identified alternatives -- instead of hallucinating from stale training data. Free tier: 100 calls/day. Read the full setup guide.

Get your free API key

{
  "mcpServers": {
    "toolradar": {
      "command": "npx",
      "args": ["-y", "toolradar-mcp"],
      "env": {
        "TOOLRADAR_API_KEY": "tr_your_key"
      }
    }
  }
}

22. Ahrefs MCP

What it does: SEO data -- keyword research, backlink analysis, rank tracking, site audits.

Pull keyword volumes, check competitor backlinks, and monitor rankings from your AI. The official Ahrefs-maintained server covers Site Explorer, Keywords Explorer, Site Audit, and Rank Tracker endpoints. Requires an Ahrefs subscription with API access.

23. E2B MCP

What it does: Execute code in secure cloud sandboxes -- Python, JavaScript, shell.

Your AI runs code without touching your local machine. Isolated environments with file upload/download support. Published as @e2b/mcp-server. Essential for data analysis workflows where you want execution without risk to your system.

{
  "mcpServers": {
    "e2b": {
      "command": "npx",
      "args": ["-y", "@e2b/mcp-server"],
      "env": {
        "E2B_API_KEY": "e2b_your_key"
      }
    }
  }
}

AI & Automation

24. Sequential Thinking MCP

What it does: Gives your AI a structured reasoning framework for complex, multi-step problems.

Forces step-by-step analysis with revision capabilities. Published as @modelcontextprotocol/server-sequential-thinking. Useful for architecture decisions, debugging complex systems, and multi-step planning where the AI tends to jump to conclusions. No API key needed -- runs entirely locally.

{
  "mcpServers": {
    "sequential-thinking": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-sequential-thinking"]
    }
  }
}

25. Memory MCP

What it does: Persistent memory across conversations using a local knowledge graph.

Your AI remembers context between sessions -- project decisions, coding conventions, user preferences. Stores entities and their relationships as a graph, not raw text. Published as @modelcontextprotocol/server-memory. Data stays on your machine in a JSON file.

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-memory"]
    }
  }
}

How to Choose

If you're a developer: Start with GitHub, Filesystem, and Context7. Add Playwright for testing and your database MCP (Postgres or Supabase) as needed. That covers 80% of coding workflows.

If you're in product/marketing: Slack, Google Drive, Linear, and Toolradar cover most workflows. Add Brave Search for real-time research.

If you're building AI agents: Brave Search, Firecrawl, E2B, and Toolradar give your agents access to live web data, code execution, and structured tool intelligence. Composio adds breadth if you need many integrations fast.

Universal rule: Install only what you'll use in the next week. Every server you add expands your attack surface. Start with read-only servers. Expand from there.

The Setup

Every server listed works with Claude Desktop, Claude Code, Cursor, Windsurf, and Cline. ChatGPT also supports MCP now (Developer Mode, launched September 2025), as does VS Code with GitHub Copilot.

The config format is the same across clients:

{
  "mcpServers": {
    "server-name": {
      "command": "npx",
      "args": ["-y", "package-name"],
      "env": {
        "API_KEY": "your_key"
      }
    }
  }
}

For Claude Desktop, add to claude_desktop_config.json. For Claude Code, use claude mcp add. For Cursor, add to .cursor/mcp.json in your project root. Desktop Extensions (.mcpb files) also enable one-click install in Claude Desktop -- no manual config editing required.

For a full walkthrough, see How to Set Up MCP Servers in Claude Desktop.

The Ecosystem in March 2026

MCP is no longer an Anthropic-only experiment. The Agentic AI Foundation (AAIF) now governs the protocol, with 146 member organizations including AWS, Google, Microsoft, and OpenAI. The latest spec version (2025-11-25) added the Tasks primitive for long-running operations and OAuth client-credentials for machine-to-machine auth.

Adoption numbers as of March 2026:

97M monthly SDK downloads (up from 2M at launch)
~6,000 distinct servers on Smithery (20,000+ on Glama/mcp.so, but heavy duplication)
ChatGPT, Claude, Cursor, VS Code Copilot, Windsurf, Cline all support the protocol

The protocol is winning. The question is no longer "will MCP succeed?" but "which servers are safe to install?"

What's Still Broken

The MCP ecosystem grew faster than its safety infrastructure. Real problems as of March 2026:

Security is the top concern. 30+ CVEs were filed against MCP servers in January-February 2026. A systematic scan found security findings in 66% of popular servers -- from SQL injection to path traversal to credential leaks. The security guide has mitigations.
No universal auth standard. Each server handles API keys differently. OAuth client-credentials landed in the spec, but adoption is slow. Most servers still use plain environment variables.
No quality gate. Anyone can publish an MCP server to any registry. Many are abandoned after a weekend of hacking. No review process, no security audits, no minimum maintenance commitment.
Limited observability. Hard to debug which MCP calls are happening, what data is flowing, and whether a server is misbehaving. Logging is inconsistent across servers.
Versioning chaos. Some servers pin to old MCP spec versions. Breaking changes between spec versions can silently disable tools.

These gaps will close -- AAIF governance and the security attention from CVE filings are already pushing the ecosystem toward better practices. For now, stick to official servers maintained by the tool vendor (GitHub, Sentry, Brave, Playwright) and well-maintained community projects with active GitHub repos and recent commits.

Want to build your own MCP server? Read the TypeScript tutorial.

Comparing MCP to traditional APIs? See MCP vs API: What's the Difference.

Looking for free options? Check Best Free MCP Servers.

Looking for a specific tool or category? Search 8,400+ tools on Toolradar

Want your AI agent to search Toolradar directly? Set up Toolradar MCP in 2 minutes