Model Context Protocol
Best MCP Servers: A Curated, Categorized Directory
Covers the wider MCP server ecosystem across tools, categorized by what each one does and who maintains it, with a security note before you install anything.
An MCP server is a small program that connects an AI assistant to something specific, a codebase, a database, a SaaS product, a browser, so the assistant can actually read and act on real data instead of just talking about it in the abstract. The ecosystem has grown fast enough that picking a good one for a given job isn't obvious anymore, and not every listing you'll find online distinguishes a well-maintained, vendor-backed server from an abandoned side project with the same name. This is a curated, categorized list organized by what each server actually does and who stands behind it, closing with a security note worth reading before you install anything that runs code on your behalf.
The best MCP servers to start with are the official ones: Anthropic's own small reference servers for filesystem, git, and web fetch access, and vendor-built servers from the companies behind the tools themselves, like GitHub's, Slack's, Stripe's, and Figma's own MCP servers. Beyond those, well-regarded developer-tool servers like Playwright for browser automation, and Cloudflare's and Supabase's for infrastructure, cover common technical needs. A smaller number of community servers, like Context7 for current library documentation, are worth knowing too, but every MCP server runs real code, so an official, vendor-maintained server beats finding the single best name on a list.
Published
Free lesson · no account neededMeet ClaudeWhy Claude?6 minStart the free lessonKey Takeaways
- Several of Anthropic's original reference servers, including GitHub, Postgres, Slack, and Brave Search, were archived in 2026 and replaced by vendor-maintained servers from the companies themselves.
- An official server built by the vendor whose product it connects to, like GitHub's own MCP server, is generally the safer and better-maintained starting point over a same-named community alternative.
- Anthropic's remaining reference servers, like filesystem, git, and fetch, are explicitly documented as educational examples, not production-hardened software, by Anthropic's own repository.
- Developer-tool servers for browser automation, like Playwright's and Chrome DevTools', and infrastructure servers from Cloudflare and Supabase, cover some of the most common technical use cases.
- An MCP server is a real, locally running program with genuine code execution; the MCP specification itself says a tool's own description should be treated as untrusted content.
- Prefer an official, vendor-maintained server over a community one when both exist, and read what a server actually does, especially anything in a bin folder or hook, before installing it.
Copy and Paste
Prompts You Can Use Today
Audit an MCP server before installing it
I'm about to install this MCP server into my setup. Here's its repository readme and its tool definitions, pasted below. Tell me plainly who actually maintains it, whether it's an official vendor-built server or a community project, what access each tool it defines would actually grant once installed, and anything in its description or configuration that looks unusual or worth double-checking before I trust it. Content: [paste the readme and tool list].
Running a new server's own documentation through this kind of review before installing it is a reasonable habit, especially for anything that isn't one of the well-known vendor-official options.
Decide which MCP server actually fits a specific task
I need to connect an AI coding assistant to [describe the specific product or data source, for example a specific database, a specific SaaS tool, or browser automation]. Based on what you know about the current MCP server ecosystem, which server would be the sensible starting point, is there an official, vendor-maintained option for it, and what would I lose by picking a community alternative instead if one exists?
Naming the exact product or task rather than asking generically for "a good MCP server" gets a much more useful, specific answer.
Scope down a server's access before trusting it with real data
Before I connect this MCP server to anything containing real or sensitive data, help me figure out whether it supports a read-only mode, a least-privilege configuration, or scoped credentials instead of full access by default, based on its documentation below, and whether there's a safer way to test it on throwaway data first before pointing it at anything that matters. Documentation: [paste relevant setup instructions].
Many official servers, including GitHub's own, document a reduced-access flag or mode; it's worth confirming whether the one you're using has an equivalent before granting it full access by habit.
Official & First-Party
Anthropic's Own Reference Servers
A small set of servers Anthropic maintains directly, meant as a starting point and a reference implementation rather than a complete catalog.
Filesystem
Secure local file read and write access with configurable boundaries on what it's allowed to touch. One of the most commonly installed reference servers for day-to-day coding work.
Git
Reads, searches, and manipulates a git repository directly, useful for letting an assistant inspect history or stage changes without shelling out to raw git commands itself.
Fetch
Fetches a web page and converts it into a form suitable for an assistant to actually read and reason about, rather than raw, messy HTML.
Memory
A knowledge-graph-based persistent memory server, useful for giving an assistant continuity across sessions instead of starting from nothing every time.
Sequential Thinking
A structured, step-by-step reasoning tool meant to help an assistant work through a problem methodically rather than jumping straight to an answer.
A Thing Worth Knowing
Several Reference Servers Were Retired in 2026
Anyone reading an older write-up about MCP servers should know this: Anthropic's reference-servers repository archived its GitHub, PostgreSQL, Slack, Sentry, Brave Search, SQLite, Google Drive, Redis, GitLab, and Puppeteer servers, moving them to a separate archive and replacing each with a vendor-maintained server instead. A tutorial or blog post written before that change will point you at a package that no longer represents the current, actively supported option.
The logic behind the change is straightforward: a server built and maintained by the actual company behind a product, say GitHub's own MCP server rather than a generic community one with the same name, tends to stay current with that product's own API changes and gets real support behind it. The categories below reflect the servers that replaced the archived ones, not the originals, so check which one you're actually installing before trusting an older guide's exact package name.
Vendor-Official
First-Party Servers From the SaaS Companies Themselves
Built and maintained directly by the vendor whose product they connect to.
GitHub MCP Server
Repository search, issues, pull requests, code context, and GitHub Actions, built and maintained by GitHub directly. The replacement for Anthropic's archived GitHub reference server.
Slack MCP Server
Search and read or send messages, canvases, and files inside Slack, built by Slack itself as the official successor to the archived reference server.
Google Drive MCP Server
File access and search inside Google Drive, documented directly by Google Workspace as an official server for that specific product.
Sentry MCP
Query issues, traces, and errors from Sentry, built and maintained by Sentry, with a hosted version available in addition to self-hosting it yourself.
Stripe MCP
Customers, payments, subscriptions, refunds, and invoices, built and maintained by Stripe directly, with a hosted option alongside the self-hosted package.
Figma Dev Mode MCP Server
Pulls design-system context and code straight out of a Figma file into a coding tool, built and maintained by Figma itself.
Developer Tools & Infrastructure
Official Servers for Common Technical Work
Vendor-maintained servers covering browser automation and infrastructure, two of the most common things people reach for an MCP server to do.
Playwright MCP
Browser automation, page inspection, and screenshots, built and maintained by Microsoft, the same company behind the underlying Playwright project.
Chrome DevTools MCP
Inspects and controls a live Chrome instance directly, built and maintained by the Chrome team at Google.
Cloudflare MCP Server
Deploy and configure Cloudflare Workers, KV storage, R2 object storage, and D1 databases, built and maintained by Cloudflare.
Supabase MCP Server
Manages a Supabase project's database schema, authentication, storage, and raw SQL, labeled official on its own project listing and maintained by the Supabase community organization.
Brave Search MCP Server
Web and local search through Brave's own search API, built by Brave as the vendor-maintained replacement for Anthropic's archived Brave Search reference server.
Worth Knowing About
A Couple of Community Servers That Have Earned Their Place
Most of what's worth recommending in the MCP ecosystem today is vendor-official, but a small number of community-built servers are genuinely well-regarded and worth knowing specifically rather than browsing a long, unfiltered directory. Context7, built by Upstash, injects up-to-date, version-specific library documentation and code examples directly into a coding assistant's context, which solves a real and common problem: a model's own training data goes stale, and a library's API can change in ways the model simply doesn't know about yet.
A community-maintained fork of Slack's archived reference server has also picked up where that original left off, for anyone who specifically wants that older implementation's behavior rather than Slack's own newer official server. Beyond names like these with a clear, checkable maintenance history, treat a generic "awesome MCP servers" list entry with real skepticism: existing on a list is not the same thing as being actively maintained, reviewed, or safe to run without checking it yourself first, and a directory site's own star count or ranking is a popularity signal, not a safety or quality one on its own.
Before You Install Anything
A Security Note on MCP Servers
An MCP server is a real program with real code execution, and the ecosystem's own specification and several security researchers are direct about the risk.
Tool descriptions are untrusted by design
The MCP specification itself states that tools represent arbitrary code execution and that a tool's description should be treated as untrusted content unless it comes from a server you already trust.
Anthropic's own reference servers aren't production-hardened
Anthropic's reference-servers repository says directly that they're educational examples, not production-ready software, and that developers should evaluate their own security needs and add their own safeguards.
A tool's description can hide an attack
Security researchers have documented "tool poisoning," where malicious instructions are hidden inside a tool's own description or metadata, invisible to you but visible to the model, including cases where a previously approved tool's behavior changes later without obvious warning.
Treat a new server like a new dependency
Installing an unreviewed third-party MCP server is functionally the same as adding an unreviewed code dependency to something you run. Prefer an official, vendor-maintained server, use a read-only or least-privilege option when one is offered, and re-check a server's tool descriptions after any update.
Reading a New Tool's Permissions Before You Trust It Is a Real Skill
Knowing what a tool, a plugin, or an MCP server is actually allowed to do before you install it is one of the most practical habits in working with AI day to day. Nightschool AI's free lesson on setting up a Claude Code project builds that habit from the start.
Frequently Asked Questions
What is an MCP server, in plain terms?
A small program that connects an AI assistant to something specific, like a codebase, a database, or a SaaS product, so the assistant can read and act on real, current data rather than just discussing the topic abstractly from what it already happens to know.
Should I use an official server or a community one if both connect to the same product?
Prefer the official one, built and maintained by the company behind that product, when it exists. It's more likely to stay current with that product's own changes and to have real support behind it.
Why did some of Anthropic's original reference servers disappear?
Anthropic archived several of them, including GitHub, PostgreSQL, Slack, Sentry, Brave Search, and others, in 2026, replacing each with a vendor-maintained server from the company behind that product instead.
Are MCP servers safe to install?
Treat any MCP server as real code that runs with genuine access to whatever it's connected to. The protocol's own specification says a tool's description should be treated as untrusted content, and Anthropic's own reference servers are explicitly documented as educational, not production-hardened.
What does "tool poisoning" mean?
A documented attack where malicious instructions are hidden inside an MCP tool's own description or metadata, invisible to the person using it but visible to the model reading it, sometimes changing after a tool was already approved.
What's a good MCP server for keeping library documentation current?
Context7, a community-built server from Upstash, injects up-to-date, version-specific library documentation and examples directly into a coding assistant's context, which helps when a library's API has changed since the model's own training data.
Is this the same as the site's existing Claude MCP servers guide?
No. That guide walks through what MCP is and how to install and configure it specifically for Claude, with a short list of example servers used mainly to illustrate the install steps. This page is a categorized directory across the wider ecosystem, organized by what each server does and who maintains it, with a dedicated section on security before you install anything.
What's the single most important thing to check before installing any MCP server?
Who actually maintains it, an official vendor versus an unreviewed community project, and what it will actually be able to do once installed, especially anything that runs its own code outside a normal permission sandbox, like a bundled hook, a helper script, or a broad default credential scope you didn't explicitly ask for.
Build the Habit of Checking Before You Trust
The same instinct that makes any new AI tool, server, or plugin worth using safely.