Best MCP Servers in 2026: 20 Picks With Install Commands

MCP servers are what turn a chat model into an agent that can actually do things: read your GitHub issues, query a database, drive a browser, search the web or edit a Notion page. There are now thousands of them, and most “best MCP servers” lists mix official servers with abandoned weekend projects. In this guide I list the MCP servers that are worth connecting in November 2026, grouped by job, with the official endpoint or install command for each one. I also show how to add them to Claude Code, Claude Desktop, Cursor, VS Code and Cline, which ones work fully offline with a local LLM, and what the new 2026-07-28 MCP specification changed.

TL;DR

Start with these five: GitHub, Context7 (current library docs), Playwright (browser), Filesystem and one web search server (Exa, Tavily or Firecrawl).
Prefer official servers. GitHub, Notion, Linear, Slack, Sentry, Supabase, Figma, Atlassian, Hugging Face and others now host their own remote MCP endpoints with OAuth, so you do not need to run anything locally.
For local LLMs and offline use: Filesystem, Git, Memory, Fetch, Playwright, Serena and Postgres MCP Pro run on your machine over stdio.
The spec changed: the 2026-07-28 revision made MCP stateless (no more initialize handshake or session IDs), added server/discover, and deprecated the old HTTP+SSE transport, Roots, Sampling and Logging. Old /sse URLs are being shut down, so use the /mcp endpoints listed below.
Security: every server you add can read data and take actions. Use read-only modes where offered and only enable the tools you need.

What Is an MCP Server?

The Model Context Protocol (MCP) is an open standard, introduced by Anthropic at the end of 2024 and now run as a Linux Foundation project under the Apache-2.0 license, for connecting AI apps to tools and data. An MCP client (Claude Code, Claude Desktop, Cursor, VS Code, Cline, ChatGPT and many more) talks to an MCP server, and the server exposes tools (actions), resources (data) and prompts.

There are two kinds of servers:

Local servers run on your computer and talk to the client over stdio. You start them with a command such as npx, uvx or Docker. Good for files, Git, local databases and anything private.
Remote servers are hosted HTTPS endpoints (the Streamable HTTP transport), usually with OAuth sign-in. Nothing to install; you just paste a URL.

Official SDKs now exist for nine languages: Python, TypeScript, Java, Kotlin, C#, Go, Rust, Swift and Ruby, and an official MCP Registry has been in preview since September 2025.

Quick List: Best MCP Servers by Category

Category	Server	Maintainer	Type
Coding	GitHub	Official (GitHub)	Remote + local Docker
Coding	GitLab	Official (GitLab)	Remote
Coding	Context7	Official (Upstash)	Remote + local
Coding	DeepWiki	Official (Cognition/Devin)	Remote, free
Coding	Serena	Community	Local
Browser	Playwright MCP	Official (Microsoft)	Local
Search	Exa, Tavily	Official	Remote
Scraping	Firecrawl, Apify	Official	Remote + local
Database	Supabase, Neon	Official	Remote
Database	Postgres MCP Pro	Community	Local
Data	dbt	Official	Remote + local
Productivity	Notion, Linear, Slack, Asana, Atlassian	Official	Remote
Design	Figma, Canva, Miro	Official	Remote
Observability	Sentry, Grafana	Official	Remote / local
AI / ML	Hugging Face	Official	Remote
Automation	Zapier	Official	Remote
Local basics	Filesystem, Git, Memory, Fetch	Official (MCP project)	Local

Best MCP Servers for Coding

1. GitHub MCP Server

The most useful server for developers. It reads code, manages pull requests and issues, checks GitHub Actions runs and security alerts, and lets you switch toolsets on or off so the model does not see 100 tools it does not need. Use the hosted endpoint https://api.githubcopilot.com/mcp/ or run it locally with the Docker image ghcr.io/github/github-mcp-server.

2. GitLab MCP

GitLab now ships a built-in server at https://gitlab.com/api/v4/mcp (self-hosted: your instance URL plus /api/v4/mcp). An admin has to enable it, and it uses OAuth with the mcp scope or a fine-grained token.

3. Context7

Context7 feeds the model current, version-specific documentation and code examples for libraries, which stops it from inventing APIs that changed two versions ago. Remote: https://mcp.context7.com/mcp. Local: npx -y @upstash/context7-mcp. It works without a key; a free key raises the rate limit.

4. DeepWiki

Ask questions about any public GitHub repository. Free, no sign-in, remote at https://mcp.deepwiki.com/mcp.

5. Serena

Serena gives agents symbol-level code tools (find symbol, find references, replace a function body) instead of plain text search, which saves tokens on big codebases. Install with uv tool install -p 3.13 serena-agent and run serena init. Pairs well with local models, because they benefit most from precise tools.

Browser, Web Search and Scraping

6. Playwright MCP

Microsoft’s Playwright server drives a real browser using the accessibility tree, so the model can click, type and read pages without screenshots. Run it with npx @playwright/mcp@latest. Useful flags: --headless, --isolated (fresh profile) and --caps=vision. It is not a security sandbox, so do not point it at sites where you are logged in to sensitive accounts.

7. Exa and 8. Tavily

Search APIs built for agents. Exa (https://mcp.exa.ai/mcp) has web_search_exa and web_fetch_exa tools and a free rate-limited tier without a key. Tavily (https://mcp.tavily.com/mcp/) does search plus extraction.

9. Firecrawl and 10. Apify

Firecrawl turns pages into clean Markdown and crawls whole sites (remote https://mcp.firecrawl.dev/v2/mcp or local npx -y firecrawl-mcp with FIRECRAWL_API_KEY). Apify (https://mcp.apify.com) runs ready-made scrapers from the Apify Store.

Databases and Data

11. Supabase

Manage Supabase projects, run SQL, apply migrations, read logs and generate types. Remote at https://mcp.supabase.com/mcp with OAuth. Add ?read_only=true and project_ref= to limit it to one project in read-only mode, which I strongly recommend.

12. Neon

Neon’s official server manages Neon Postgres and runs SQL, aimed at local development and IDE use.

13. Postgres MCP Pro

A community server for any Postgres database with health checks, query plan review and index recommendations. Install with pipx install postgres-mcp or Docker, and use its restricted read-only mode for production databases.

14. dbt

Exposes dbt models, lineage and Semantic Layer metrics, and can run build and test locally. uvx dbt-mcp.

Productivity, Design and Observability

Notion (https://mcp.notion.com/mcp): search, read and write pages and databases.
Linear (https://mcp.linear.app/mcp, or /mcp/readonly): issues, projects and comments. The old /sse URL is deprecated.
Slack (https://mcp.slack.com/mcp): search messages and channels, send messages, create canvases.
Atlassian Rovo (https://mcp.atlassian.com/v2/mcp): Jira, Confluence, Jira Service Management and Bitbucket.
Asana (https://mcp.asana.com/v2/mcp): tasks and projects.
Figma (https://mcp.figma.com/mcp): read frames, variables and components, with canvas writing in beta. Canva and Miro have official OAuth connectors too.
Sentry (https://mcp.sentry.dev/mcp): pull issues, stack traces and events into your coding agent.
Grafana (uvx mcp-grafana): search dashboards, query Prometheus and Loki, read alerts. Needs GRAFANA_URL and a service account token.
Hugging Face (https://huggingface.co/mcp): search models, datasets, Spaces and papers, and call MCP-enabled Gradio Spaces.
Zapier (create your server at mcp.zapier.com): a fallback for any app without its own server. Turn on only the actions you need.

Best MCP Servers for Local LLMs (Offline)

If you run a model in Ollama, LM Studio or llama.cpp and want everything to stay on your PC, these servers need no cloud account:

Filesystem, Git, Memory (a local knowledge graph that remembers facts between sessions) and Fetch (turns a web page into model-friendly text). All four are reference servers in the modelcontextprotocol/servers repository.
Playwright MCP for browser automation.
Serena for symbol-level code editing.
Postgres MCP Pro or a SQLite server for local databases.
Blender MCP for 3D work. I wrote a full guide on Blender MCP with a local LLM in Cline.
Comfy MCP for ComfyUI workflows, covered in my Comfy Agent and Comfy MCP guide.

One practical tip: small local models get confused when they see dozens of tools. Connect two or three servers at a time, and pick a model that is good at tool calling. My best Ollama models list marks which ones handle agents well.

How to Add an MCP Server (Claude Code, Claude Desktop, Cursor, VS Code, Cline)

Claude Code

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

# Local stdio server
claude mcp add context7 -- npx -y @upstash/context7-mcp

# Local server with an API key
claude mcp add firecrawl -e FIRECRAWL_API_KEY=fc-your-key -- npx -y firecrawl-mcp

# Check what is connected
claude mcp list

Use --scope user to make a server available in every project, or --scope project to share it with your team through a .mcp.json file in the repo.

Claude Desktop, Cursor and Cline (JSON config)

These clients use the same mcpServers shape. In Claude Desktop it lives in claude_desktop_config.json (Settings, Developer, Edit Config), in Cursor in ~/.cursor/mcp.json (global) or .cursor/mcp.json (project), and in Cline under MCP Servers, Configure.

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": ["@playwright/mcp@latest", "--isolated"]
    },
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/path/to/your/project"]
    },
    "context7": {
      "url": "https://mcp.context7.com/mcp"
    }
  }
}

VS Code (GitHub Copilot agent mode)

VS Code reads .vscode/mcp.json in your workspace. Note the top-level key is servers, not mcpServers:

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

After saving, restart the client (or reload the MCP list) and approve the OAuth sign-in for remote servers.

What Changed in the 2026-07-28 MCP Spec

The MCP specification got its biggest revision so far on July 28, 2026. You do not need to read it to use servers, but it explains why some old setups stopped working:

MCP is now stateless. The initialize handshake and the Mcp-Session-Id header are gone. Every request carries its protocol version and client capabilities, which makes remote servers much easier to scale.
New server/discover call that every server must implement, so clients can check versions and capabilities up front.
Multi Round-Trip Requests replace server-initiated requests for things like elicitation (asking the user for input).
Tasks moved to an official extension (io.modelcontextprotocol/tasks) for long-running jobs.
Deprecated: the old HTTP+SSE transport, plus the Roots, Sampling and Logging features, and Dynamic Client Registration in favour of Client ID Metadata Documents. Deprecated features keep working for at least twelve months.
Caching hints (ttlMs, cacheScope) on list results and a recommendation to return tools in a fixed order, which improves prompt cache hit rates.

In practice: if a server’s docs still give you an /sse URL, look for the newer /mcp endpoint.

Security Checklist

Use official servers where they exist, and check who maintains a community server before running it.
Turn on read-only modes (Supabase read_only=true, Linear /mcp/readonly, Postgres MCP Pro restricted mode).
Limit scope: give Filesystem only the folders it needs, scope GitHub tokens to specific repos, and enable only the Zapier actions you use.
Watch for prompt injection. A web page, issue or document the agent reads can contain instructions. Keep human approval on for actions that write or send data.
Never paste API keys into prompts. Put them in environment variables in the server config.

MCP Server Not Working? Fixes

Symptom	Likely cause	Fix
Server shows as failed or disconnected	Node.js missing or too old, or `npx` not on PATH	Install Node.js LTS, then run the same `npx` command in a terminal to see the real error
Works in terminal, fails in Claude Desktop	The app cannot see your shell PATH	Use the full path to `npx` or `node` in the config (find it with `which npx` or `where npx`)
Config changes ignored	Client not restarted or JSON is invalid	Fully quit and reopen the app; check the JSON with a validator (a trailing comma is the usual culprit)
Remote server keeps asking to sign in	OAuth token expired or blocked pop-up	Remove and re-add the server, then complete the browser sign-in
Tools do not appear in the chat	Too many servers or tools disabled	Check the tools menu or `/mcp` in Claude Code; turn off servers you do not need, since every tool uses context
VS Code ignores the file	Wrong key	VS Code uses `servers`, not `mcpServers`, in `.vscode/mcp.json`
Windows: “spawn npx ENOENT”	Windows needs the cmd wrapper	Use `"command": "cmd"` with `"args": ["/c", "npx", "-y", "package-name"]`

Where to find logs: in Claude Code run claude --debug; in Claude Desktop open Settings, Developer and view the MCP logs; in Cursor open the Output panel and choose MCP.

Starter Sets: Which Servers to Install First

You are…	Install these
A web developer	GitHub, Context7, Playwright, your database (Supabase, Neon or Postgres MCP Pro)
A researcher or writer	Exa or Tavily for search, Firecrawl for reading pages, Notion or your notes app
A data analyst	Postgres MCP Pro or dbt, plus a filesystem server limited to your data folder
Running local LLMs offline	Filesystem, a local search server and a SQLite server; keep the tool count low for small models

Start with three or four servers. Each server adds tool descriptions to every prompt, which uses context and can confuse smaller models.

Frequently Asked Questions

What is the most useful MCP server?

For developers, the GitHub MCP server and Context7. For everyone else, a web search server (Exa or Tavily) plus the official server for the app you use most, such as Notion or Slack.

Are MCP servers free?

Most servers are free and open source. Remote servers from SaaS companies are free to connect but use your existing account and plan. Search and scraping servers (Exa, Tavily, Firecrawl, Apify) have free tiers and paid limits.

Can I use MCP servers with a local LLM?

Yes. Cline, LM Studio, Open WebUI and other clients connect local models to MCP servers. Filesystem, Git, Memory, Fetch and Playwright all run offline. Use a model with strong tool calling and keep the number of tools small.

Does ChatGPT support MCP?

Yes. OpenAI and Google both support MCP in their agent platforms, so the same remote servers work across Claude, ChatGPT, Gemini-based tools, Cursor and VS Code.

Where can I find more MCP servers?

The official MCP Registry, the modelcontextprotocol/servers GitHub repository, mcp.so and the awesome-mcp-servers list. Prefer servers published by the company that owns the service.

Final Thoughts

You do not need fifty MCP servers. Start with GitHub, Context7, Playwright, Filesystem and one search server, then add the official server for each app you use every day. Prefer remote official endpoints for SaaS tools and local stdio servers for anything private. If you are choosing which coding agent to plug them into, read my Cursor vs Windsurf vs Claude Code vs Cline comparison, and if you want the whole stack offline, set up Ollama first.

Best MCP Servers in 2026: 20 Picks for Claude, Cursor, VS Code and Local LLMs (With Install Commands)