The Best MCP Servers for Claude Code in 2026 (Most Lists Are Recommending Dead Ones)

Short answer: the official Model Context Protocol repository actively maintains exactly seven reference servers: Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking and Time. The ones most roundups still recommend, GitHub, PostgreSQL, Puppeteer, Slack, Sentry, SQLite, Redis, Google Drive and Google Maps, have all been moved to an archived repository. Some have official replacements maintained elsewhere. Install from the maintained set first, then add third-party servers deliberately.

If you have searched for MCP servers recently you have probably read a list of fifteen or twenty, presented as a menu of equals. That list is almost certainly out of date, and not in a subtle way. A large share of the servers people recommend are no longer maintained in the official repository at all.

This matters more than a stale link. An archived server still installs. It still runs. It just stops receiving fixes, and you find out months later when something breaks and nobody is coming.

What is actually maintained

These are the reference servers in the official modelcontextprotocol/servers repository today:

Seven. That is the maintained reference set. Everything else you install comes from somewhere else, which is fine, as long as you know that is the trade you are making.

What moved to the archive

These were reference servers and are now in servers-archived:

AWS KB Retrieval, Brave Search, EverArt, GitHub, GitLab, Google Drive, Google Maps, PostgreSQL, Puppeteer, Redis, Sentry, Slack and SQLite.

Read that list again, because it is most of what a typical "best MCP servers" article recommends. GitHub and Postgres in particular show up in nearly every roundup as though they were current.

Two of them have official homes elsewhere rather than simply being abandoned. Brave Search is now maintained by Brave. Slack is now maintained by Zencoder. Those are real, supported options. The difference is that you are now trusting a different maintainer, and you should know that before you wire it into your daily workflow.

How to choose, in practice

Start with Filesystem and Git. Between them they cover the majority of what Claude Code does on a normal day: read the project, understand history, make changes. If you install nothing else, install these two.

Add Fetch when you need current information. A model working from training data will confidently tell you about a library version that no longer exists. Fetch lets it read the actual page.

Add Memory if your work spans sessions. It is a knowledge graph, not a chat log, which means it is good at facts that stay true and poor at remembering a conversation.

Be honest about Sequential Thinking. It structures multi-step reasoning. On a genuinely hard problem it helps. On a simple one it adds latency and tokens for no gain.

Treat everything else as a third-party dependency, because that is what it is. Check when it was last updated, check who maintains it, and check what credentials it wants.

The permission question nobody asks first

An MCP server runs with whatever access you hand it. A filesystem server can read any path you allow. A database server can read any table it can reach.

Before installing one, ask what happens if it is compromised or simply buggy. Scope the filesystem server to your project directory rather than your home directory. Give database servers a read-only user. Prefer servers that ask for narrow credentials over ones that want an admin token, and be suspicious of any that will not tell you which permissions they need.

This is ordinary supply-chain hygiene. It is worth stating plainly because the install instructions rarely mention it.

Why so many lists are wrong

MCP moved fast, and content about it was written faster. An article published when GitHub and Postgres were reference servers was accurate on the day it went out. Nothing updated it when the repository was reorganised, and it kept ranking.

The fix is not to trust any list, including this one, over the source. Open the official servers repository and look at what is in the main tree versus the archive. It takes a minute and it is the only version that is current by definition.

FAQ

How many official MCP servers are there? Seven reference servers are actively maintained: Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking and Time. Many more exist as third-party or company-maintained servers.

Is the GitHub MCP server dead? The version that lived in the official reference repository is archived. GitHub-related functionality is available from other maintainers, so check the source and the last commit date before relying on it.

Which MCP servers should I install first? Filesystem and Git. They cover most day-to-day work with Claude Code. Add Fetch when you need the model to read current pages.

Do MCP servers slow Claude Code down? Each connected server adds tool definitions to the context and some latency. A small, deliberate set beats installing everything you find.

Are archived servers unsafe to use? Not automatically. They are unmaintained, which means known issues will not be fixed. That is an acceptable risk for a scratch project and a poor one for anything handling credentials or production data.