The Best MCP Servers in 2026: A Curated List for Real Work

Why a curated list beats a giant directory

Search "MCP servers" today and you'll find thousands of them: official reference implementations, vendor-maintained integrations, and a long tail of weekend projects that haven't been touched since early 2025. The community's answer to that sprawl is awesome-mcp-servers, a GitHub list with tens of thousands of stars that catalogs the ecosystem — useful as a map, but a map isn't a recommendation. It tells you what exists, not what's worth trusting with your filesystem, your GitHub token, or your production database.

This post is the opposite of that list. It's a short, opinionated set of the best MCP servers we'd actually recommend installing in 2026, organized by category, with a one-line reason for each pick and a link to its real source. Every server below is either an official reference implementation from the Model Context Protocol project, or built and maintained by the company whose product it connects to. We skipped anything we couldn't verify was actively maintained. Before you install anything from this list or any other, read our guide to vetting MCP servers before you trust them — installing a community MCP server is closer to installing an unaudited dependency than calling a REST API.

If you're new to the protocol itself, start with our MCP primer first; this post assumes you already know what MCP is and just want good servers to install.

How we picked these

Four criteria, applied consistently:

1. Real maintainer. The repo is owned by the company that runs the underlying service (GitHub, Sentry, Neon, Notion) or by the Model Context Protocol organization itself — not a third-party reimplementation of someone else's API. This matters more for MCP than for a typical library dependency: an MCP server usually runs with your credentials attached, so the further it drifts from the company that owns the underlying system, the harder it is to reason about what it's doing with them.
2. Active in 2026. Commits, releases, or changelog entries within the last few months, not an abandoned proof of concept from MCP's 2024 launch window. The protocol itself has moved fast — transports, auth, and tool-annotation conventions have all shifted since the early servers were written — so a server that hasn't been touched in a year is often quietly running against an older spec revision.
3. Scoped, legible tools. The server exposes a manageable set of tools with clear names and descriptions, not a 200-tool kitchen sink that blows out your context window. Tool descriptions double as instructions an agent reads and acts on, so vague or overly broad tool surfaces are also a security smell, not just a usability one.
4. Solves a real workflow. Each pick earns its place by doing something you'd actually reach for while coding, researching, or running an agent day to day, not a novelty demo that looks good in a launch tweet but doesn't survive contact with a real project.

We deliberately left a lot out. Several categories — Slack, Airtable, Stripe, most CRMs — have multiple credible options but no single server that's clearly the default the way GitHub's or Sentry's official servers are for their categories; we'd rather point you to the full directory for those than force a pick. We also skipped anything still labeled experimental or pre-1.0 by its own maintainer, even if the underlying idea was good, because "best" should mean stable enough to build on, not just interesting.

Official reference servers

These live in modelcontextprotocol/servers, maintained directly by the MCP project. They're the safest possible starting point because there's no third party in the trust chain.

Filesystem

Secure, configurable file operations — read, write, move, and search files within directories you explicitly allow. It's the server every local coding agent setup eventually needs, and because access is scoped to allow-listed directories, it's a reasonable first server to trust. Best for: giving an agent controlled access to a local project folder.

Repo: modelcontextprotocol/servers/tree/main/src/filesystem

Fetch

Pulls web content and converts it to clean, LLM-friendly text, handling the HTML-to-markdown conversion so your model isn't parsing raw tag soup. It's deliberately minimal — one job, done well. Best for: quick, no-auth-required page reads inside an agent loop.

Repo: modelcontextprotocol/servers/tree/main/src/fetch

Git

Read, search, and manipulate local Git repositories: diff, log, branch, and commit inspection without shelling out to raw git commands from the model. Best for: code-review and repo-archaeology agents that need to reason about history, not just current files.

Repo: modelcontextprotocol/servers/tree/main/src/git

Memory

A local knowledge-graph server that gives an agent persistent memory across sessions — entities, relations, and observations stored and queried between conversations. Best for: long-running assistants that need to remember facts about a project or user without re-deriving them every session.

Repo: modelcontextprotocol/servers/tree/main/src/memory

Dev tools

GitHub MCP Server

GitHub's own official server, rewritten in Go and built jointly with Anthropic to replace the earlier community version. It covers repository browsing, issue and PR management, code search, Actions/workflow inspection, and security-alert triage, and ships both as a GitHub-hosted remote server and a local binary. Best for: agentic PR review, issue triage, and CI-failure investigation directly against your real repos.

Repo: github/github-mcp-server

Playwright MCP

Microsoft's official browser-automation server for Playwright. Instead of screenshots and vision models, it drives the browser using Playwright's accessibility tree, which makes clicks and form fills deterministic rather than guesswork. Ships as @playwright/mcp on npm and as a Docker image. Best for: agents that need to test, scrape, or operate real web UIs, not just fetch static pages.

Repo: microsoft/playwright-mcp

Sentry MCP

Sentry's official server for pulling issues, stack traces, and event details straight into a coding assistant, with natural-language issue search and state-triage tools. It runs as a hosted remote server at mcp.sentry.dev with OAuth login against your existing Sentry org, or locally over stdio. Best for: "why did this deploy break" loops where the agent needs live error data, not just code.

Repo: getsentry/sentry-mcp · Docs: docs.sentry.io/product/sentry-mcp

Context7

Upstash's documentation server that pulls current, version-specific library docs and code examples straight into your prompt, which matters a lot for fast-moving frameworks where a model's training data is already stale. Runs as a hosted MCP endpoint (mcp.context7.com/mcp) or locally via @upstash/context7-mcp. Best for: cutting down hallucinated API calls when an agent is writing code against a library it doesn't know well.

Repo: upstash/context7

Data

Neon MCP Server

Neon's official server for managing serverless Postgres databases and running queries in natural language — create projects and branches, run migrations, execute SQL, and manage the Neon API surface from an agent. The Neon team is explicit that it's meant for local development and IDE workflows, not production database access from an untrusted agent. Best for: prototyping schema changes and running one-off queries during development.

Repo: neondatabase/mcp-server-neon

Notion MCP Server

Notion's official server for searching, reading, creating, and updating pages, databases, and workspace content. Version 2.0 moved to Notion's 2025-09-03 API and its "data sources" model for databases; Notion has also signaled it's prioritizing a hosted remote MCP endpoint over this local server going forward. Best for: agents that need to read or write structured docs and databases your team already lives in.

Repo: makenotion/notion-mcp-server

Web and search

Firecrawl MCP

The official server for Firecrawl's scraping and search API — turns arbitrary web pages into clean, agent-ready markdown, with crawl, map, and structured-extract tools beyond plain fetching. A keyless free tier covers basic scrape/search; crawling and extraction need an API key. Best for: research agents that need to pull structured data from many pages, not just one URL at a time.

Repo: firecrawl/firecrawl-mcp-server

Exa MCP Server

Exa Labs' official server for their neural search API — real-time web search with content extraction, plus purpose-built tools like company research and LinkedIn search. It's built specifically for agent consumption rather than adapting a general search API after the fact. Best for: research and sourcing tasks where you want ranked, relevant results rather than raw SERP scraping.

Repo: exa-labs/exa-mcp-server

Productivity

Linear MCP

Linear's first-party remote server for reading and writing issues, projects, and comments — search and filter issues, check sprint status, surface blocked tickets, and create or update issues with labels and assignees. It authenticates via OAuth in the browser, so there's no API key to manage or leak. Best for: teams that already track work in Linear and want an agent to triage or update tickets directly.

Docs: linear.app/docs/mcp

Zapier MCP

A hosted server that puts Zapier's catalog of thousands of app integrations behind MCP tools, so an agent can trigger actions across whatever your team already has connected to Zapier without a bespoke integration for each one. It's the broadest-reach server on this list by design — useful precisely because it's shallow but wide. Best for: "send this to Slack / add this to a spreadsheet / create a calendar event" style actions without writing a dedicated integration.

Repo: zapier/zapier-mcp · Docs: docs.zapier.com/mcp

Setting up MCP servers in Claude Code

If you're adding any of these as claude code mcp servers, the CLI has a single command for it: claude mcp add. The basic shape is:

# Remote HTTP server (recommended when the vendor offers one)
claude mcp add --transport http notion https://mcp.notion.com/mcp

# Local stdio server, launched as a subprocess
claude mcp add context7 -- npx -y @upstash/context7-mcp

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

A few things worth knowing before you run it:

• Everything after -- is the actual command Claude Code spawns to launch the server; flags like --transport and -e go before it.
• The default --scope local keeps a server private to you in the current project. Use --scope user to make it available everywhere, or --scope project to commit it to .mcp.json and share it with your team.
• Prefer --transport http for any server that offers a hosted endpoint (GitHub, Sentry, Linear, Notion, Zapier all do) — you skip local process management and API-key storage entirely, and auth happens through OAuth in the browser.
• Run claude mcp list afterward to confirm the server connected, and check the tool count it reports — an unexpectedly large number of tools from one server is worth investigating before you start using it in earnest.

A practical starting stack for most Claude Code users: Filesystem and Git for local project work (often unnecessary if you're only using Claude Code's own built-in file and shell tools, but useful when you want an MCP-native equivalent), GitHub MCP Server for anything touching pull requests or issues, and Context7 so the model stops guessing at library APIs it hasn't seen recently. Add the rest — Sentry, Linear, Notion, Firecrawl — as the workflows come up, rather than installing everything on day one. Every additional server is additional context in the model's tool list and additional trust you're extending, and both have real costs even when nothing goes wrong.

The list at a glance

Server	Category	Best for	Link
Filesystem	Official reference	Scoped local file access	repo
Fetch	Official reference	Simple web page reads	repo
Git	Official reference	Repo history and diffs	repo
Memory	Official reference	Persistent agent memory	repo
GitHub MCP Server	Dev tools	PR/issue triage on real repos	repo
Playwright MCP	Dev tools	Browser automation and testing	repo
Sentry MCP	Dev tools	Live error and stack-trace triage	repo
Context7	Dev tools	Current library docs in-prompt	repo
Neon MCP Server	Data	Serverless Postgres from natural language	repo
Notion MCP Server	Data	Reading/writing team docs	repo
Firecrawl MCP	Web/search	Structured scraping across many pages	repo
Exa MCP Server	Web/search	Neural web search for research agents	repo
Linear MCP	Productivity	Issue tracking from an agent	docs
Zapier MCP	Productivity	Broad app-to-app actions	repo

Where to go from here

Fourteen servers is a starting toolbox, not the whole ecosystem — there are strong options for Slack, Airtable, Stripe, and dozens of other tools we didn't have room for here. For the full, continuously updated catalog, browse the mcpservers.org directory, where you can filter by category and check maintenance status before you install anything.

Treat this list as a floor, not a ceiling. New official servers ship every month as more companies realize an MCP integration is now table stakes alongside a REST API and a webhook, and some of today's community favorites will graduate into first-party servers the way GitHub's did. Revisit your setup periodically rather than installing once and forgetting about it — a server that was the best option when you added it can be superseded, and one that looked actively maintained a year ago can quietly stall out. And whatever you add, read it through the lens of our MCP security best practices guide first — the best MCP servers are the ones you've actually vetted, not just the ones with the most GitHub stars.