Best MCP Servers for Claude, Cursor and Codex (November 2026)

Short version: nearly two years after Anthropic published the Model Context Protocol, MCP is the plug standard for giving a coding agent access to the outside world, and it is no longer Anthropic's alone: on 9 December 2025 the protocol was donated to the Agentic AI Foundation under the Linux Foundation, co-founded with OpenAI and Block, which cited more than 10,000 published servers at the time. Claude Code, Cursor and Codex all speak it. The problem is choosing. Below are the servers we would actually install, with star counts, latest releases and caveats pulled from the GitHub API on 10 November 2026, and a note on the ones that quietly died. This page is refreshed monthly.

1. modelcontextprotocol/servers (91,899 stars)

The official reference implementations, maintained by the protocol project itself and released as v2.3.0 on 4 November. The repository no longer tries to be a directory of every server (it points you to the MCP Registry for that) and instead keeps a small set of canonical servers: filesystem, fetch, git, memory, sequential thinking and time. Start with the filesystem and fetch servers; between them they cover most of what a coding agent asks for before it needs anything exotic. The README is honest that these are reference, educational implementations rather than hardened production servers, which is fine on a workstation and a reason to read their code: it is the quickest way to learn how a well-behaved server is written.

2. upstash/context7 (64,447 stars)

Context7 fixes the single most common failure of coding agents: answering from training data about a library that has since changed. It pulls version-specific documentation and code examples into the prompt at request time, so the agent writes against the API you actually have installed. It ships several times a month (three releases in the first nine days of November, 4.6.0 being the latest) and works with Claude Code, Cursor and Codex alike; the README now also offers a hosted endpoint and an npx ctx7 setup installer if you would rather not run it locally. If you install only one third-party server, make it this one.

3. ChromeDevTools/chrome-devtools-mcp (55,049 stars)

Google's Chrome DevTools team ships this one, and it is the browser server to reach for when the question is not "click through this flow" but "why is this page slow". It lets the agent control and inspect a live Chrome with the full DevTools surface, performance traces, network requests, console output, and the README names Claude, Cursor, Copilot and Antigravity as clients. Created in September 2025, it is already past 55,000 stars; v1.12.0 shipped on 22 October and the repository is pushed daily. It also ships as a plain CLI for agents that prefer one.

4. microsoft/playwright-mcp (39,400 stars)

Microsoft's Playwright team exposes a real browser to the agent through structured accessibility snapshots rather than screenshots, which keeps it fast, cheap in tokens and usable with models that are not vision-tuned. Releases land every few weeks (v0.0.85 on 9 November, v0.0.84 on 31 October), and the README now also offers a CLI mode that it says agents increasingly prefer. Use it for anything that needs a logged-in session, a form, or a page without an API, and prefer it over screenshot-based browser servers unless you specifically need pixels.

5. github/github-mcp-server (34,359 stars)

GitHub's official server, written in Go, lets the agent browse code, search files, read and manage issues and pull requests, inspect commits and work with Actions, all with your GitHub permissions. v2.1.0 shipped on 3 November. Scope the token you give it: a fine-grained personal access token with read-only repository access is enough for most review and research tasks, the server has a --read-only flag for exactly that, and nothing else in this list is more sensitive if a prompt injection lands.

6. oraios/serena (31,257 stars)

Serena describes itself as the IDE for your coding agent, and that is fair: it uses language servers to give the agent symbol-level retrieval, editing, refactoring and debugging instead of grep and whole-file rewrites. On large codebases this is the difference between an agent that finds the right function in one call and one that burns the context window reading files. v2.1.0 came out on 19 October and the project is pushed to regularly.

7. googleapis/mcp-toolbox (16,832 stars)

Google's MCP Toolbox for Databases is the server we reach for when an agent needs real data: it sits between the agent and PostgreSQL, MySQL, BigQuery, Spanner and a long list of other engines, and it handles connection pooling and authentication so the agent never holds raw credentials. Written in Go, v1.16.0 shipped on 5 November. Tools are declared in a YAML file rather than code, and it ships client SDKs for calling them from your own applications.

8. GLips/Figma-Context-MCP (16,186 stars)

Framelink's Figma server gives Cursor, Claude Code or any MCP client structured layout data straight from a Figma file, which is dramatically better than pasting screenshots when you want a component implemented in one shot. It is the slowest-moving entry in this list: v0.13.2 is from June 2026 and the main branch has not changed since, so treat it as stable rather than active and check that it still matches Figma's current API before relying on it.

9. apify/apify-mcp-server (15,528 stars)

Apify's official server connects the agent to its marketplace of ready-made Actors for web scraping, data extraction and browser automation: social media, search engines, maps, e-commerce sites and more, through one connection. v0.21.4 shipped on 6 November. It is the pragmatic choice when you need data from a site that fights scrapers and you would rather not maintain the scraper yourself.

10. CursorTouch/Windows-MCP (11,447 stars)

Computer use for Windows: the agent sees and operates the desktop, not just a browser. v0.8.8 is from 10 October. It fills a gap that most of the ecosystem, which is Mac and Linux first, leaves open, and it is the one to try when the thing you want automated only has a Windows GUI.

11. mobile-next/mobile-mcp (10,463 stars)

Mobile MCP lets an agent drive iOS and Android apps on simulators, emulators and real devices through one platform-agnostic interface, so the same prompt can test a flow on both platforms. Release 1.0.16 shipped on 2 November. If you build mobile apps, this and Playwright MCP are the pair that make agent-written UI tests realistic.

12. wonderwhy-er/DesktopCommanderMCP (10,277 stars)

Desktop Commander turns a chat client into a terminal: it searches, edits and manages files and runs processes, including long-running ones, using your existing client subscription rather than API tokens. v0.3.1 arrived on 22 October. It is powerful and therefore dangerous; run it with a dedicated user account or inside a container if you let it near anything that matters.

Worth knowing, not in the top twelve

• awslabs/mcp (9,899 stars): AWS Labs' suite of servers for AWS services, still released continuously (the 9 November build is the latest), but its README now describes the Agent Toolkit for AWS as the successor to these servers, plugins and skills. Install it for today's work and expect to migrate.
• headroomlabs-ai/headroom (79,278 stars, v0.44.0): primarily a compression library and proxy that shrinks tool outputs, logs and RAG chunks before they reach the model, with an MCP server mode as a secondary interface. It sits in front of everything above and is the fastest-growing repo in this whole survey; it is left out of the twelve because serving MCP is not its main job.
• activepieces/activepieces (25,459 stars): a workflow automation platform that exposes roughly 400 integrations as MCP servers, useful when you need many SaaS connections at once.
• LaurieWired/GhidraMCP (11,872 stars, 2.0 on 10 October): reverse engineering with Ghidra, a niche where agents turned out to be unusually good.
• hangwin/mcp-chrome (12,538 stars): controls your real Chrome through an extension, but the last push was January 2026, so treat it as unmaintained until that changes.
• BeehiveInnovations/pal-mcp-server (11,836 stars, formerly Zen MCP): routes one prompt across several model providers. Last release December 2025.
• firecrawl/firecrawl-mcp-server (7,793 stars, v3.36.1), exa-labs/exa-mcp-server (5,220) and tavily-ai/tavily-mcp (2,483): the search and crawl trio, all actively maintained, pick by whichever API you already pay for.

Building and testing your own

If the server you need does not exist, writing one is an afternoon's work. PrefectHQ/fastmcp (28,366 stars, v4.1.7 on 10 November) is the Python framework almost everyone uses: a decorated function becomes a tool. The official python-sdk (24,820 stars, v2.3.0), which bundles FastMCP 1.0, and typescript-sdk (13,741 stars, v2.4.0) are the vendor-neutral baseline; FastMCP 2 is the standalone framework built on top of that idea. Test with modelcontextprotocol/inspector (11,268 stars, 2.14.0), which now ships web, CLI and terminal interfaces, and publish to the MCP Registry (7,455 stars), whose API has been frozen at v0.1 since October 2025; the freeze was announced as a short one ahead of a v1, but v0.1 is still the live contract clients code against. If you run servers for a team, docker/mcp-gateway packages them as containers behind one gateway, which is also the cleanest answer to the security question below.

Gone or moved since last year

• browserbase/mcp-server-browserbase was archived in July 2026 as a reference implementation; Browserbase now points users to its hosted MCP server at mcp.browserbase.com, which is built on its Stagehand framework.
• Azure/azure-mcp: development moved to microsoft/mcp in August 2025 and the old repository was formally archived in February 2026; the Azure MCP Server continues there (4.0.0 beta 6 shipped on 6 November).
• mem0ai/mem0-mcp was archived in March 2026; use Mem0's hosted MCP endpoint or its SDK instead.
• awslabs/mcp is being superseded by the Agent Toolkit for AWS, see above.

Installing one in Claude Code, Cursor and Codex

Every server's README has the exact command, but the shape is the same everywhere. Claude Code and Codex register servers from the command line; Cursor reads a JSON file in your project.

# Claude Code
claude mcp add context7 -- npx -y @upstash/context7-mcp

# Codex CLI
codex mcp add context7 -- npx -y @upstash/context7-mcp

# Cursor: .cursor/mcp.json
{
  "mcpServers": {
    "context7": { "command": "npx", "args": ["-y", "@upstash/context7-mcp"] },
    "playwright": { "command": "npx", "args": ["@playwright/mcp@latest"] }
  }
}

The security rule nobody wants to hear

An MCP server runs with whatever permissions you give it, and the agent will call it with whatever a web page or an issue comment manages to inject into the conversation. So: install official or registry-listed servers, pin versions instead of @latest for anything that touches production, hand out read-only tokens by default, and put filesystem and terminal servers in a container or a throwaway user account. The Docker MCP Gateway exists for exactly this reason.

How this list is built

Star counts, release tags, dates and archive status come straight from the GitHub API, captured on 10 November 2026. The twelve are our editorial picks from the servers that pass three filters: the project must primarily be an MCP server (frameworks, SDKs and tools with a secondary MCP mode are listed separately), must have pushed code in the last 60 days and must have shipped a release in 2026. Within those filters we favour breadth of use over raw stars, which is why GhidraMCP, a superb but niche reverse-engineering server, sits in the watch list. We re-run the numbers monthly and log every change at the top of this page. For the frameworks that consume these servers, see our AI agent repositories list.