The best MCP servers in 2026 are the maintained, first-party ones you scope tightly — not the long listicles still recommending archived projects. Start with the official reference servers for the basics (Filesystem, Git, Fetch, Memory, Sequential Thinking) and add GitHub's own server (github/github-mcp-server) for dev work, Playwright MCP for the browser, and a maintained database server such as Supabase MCP or Neon MCP when it matches your stack. The catch most guides miss: 14 legacy reference entries (GitHub, PostgreSQL, Slack, Brave Search, Google Drive, legacy Git…) are now archived — do not install them as live picks. And treat every server as attack surface: use official servers where possible, pin versions, and restrict scopes.
Which are the best MCP servers to install in 2026?
Pick by category and maintainership, not by listicle length. The Model Context Protocol (MCP) lets an AI client — Claude Desktop, Claude Code, Cursor, VS Code — call external tools through a small server process, and by mid-2026 there are thousands of them. The shortlist worth your time is small: the seven official reference servers for local primitives (files, git, web fetch, memory, reasoning), plus a handful of vendor-maintained servers for the things developers actually wire up — GitHub, a browser, a database. If you only want to understand the protocol itself, read our explainer on what MCP is and how it works first; this page assumes you already know that and want to choose servers.
One correction up front, because nearly every competing “best MCP servers” roundup gets it wrong: the GitHub, PostgreSQL, Slack, Brave Search, Google Drive and several other legacy reference servers that used to ship in the official repo are now archived. They still appear in stale lists as “the official servers,” but they are historical reference only. Installing them today means running unmaintained code. Below, every pick is checked against primary sources so you do not inherit that error.
What makes a good MCP server in 2026?
A good MCP server in 2026 is maintained by a first party, scoped to the narrowest permission it needs, and explicit about its transport and auth model. The protocol surface has matured, so the differentiator is no longer “does it have tools” — it is whether you can trust and contain it. Use this four-point rubric before adding any server to a config:
- Maintainership. First-party (the vendor that owns the API, e.g. GitHub's own server) beats a popular community wrapper. The official reference repo explicitly calls its servers educational examples, not production-ready — useful primitives, but qualify them as such.
- Transport. Local
stdioservers run as a child process on your machine; remote (hosted) servers run over HTTP and need OAuth. GitHub now offers both; the remote option removes local setup but shifts trust to the hosted service and the token scope you grant. - Auth and scope. Prefer servers that take a least-privilege token (read-only where possible) and let you limit what they touch — allowed directories, single repository, one database role.
- Security posture. Path-traversal and missing-auth defaults are common (see the security section). A good server documents its sandboxing; a bad one assumes your whole filesystem.
Where to discover servers reliably: the official MCP Registry, launched in preview on 8 September 2025 as an “app store for MCP servers.” It is the canonical discovery source, but it is still preview with no data-durability guarantee — treat it as the index, not a security stamp.
Which official MCP servers are live, and which were archived?
There are exactly seven current official reference servers, and fourteen legacy reference entries were archived. Per the official docs and the modelcontextprotocol/servers repository, the live set is: Everything (a test/demo server), Fetch (web content retrieval), Filesystem (secure file ops), Git (read/search/manipulate repos), Memory (a knowledge-graph persistent memory), Sequential Thinking (structured multi-step reasoning), and Time (time and timezone conversion). The archived list also contains a legacy Git entry, but the current Git reference remains active in the main repository.
| Status | Servers | What to do |
|---|---|---|
| Live reference (7) | Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking, Time | Use as primitives; remember they are educational, not hardened. |
| Archived legacy entries (14) | AWS KB Retrieval, Brave Search, EverArt, Git, GitHub, GitLab, Google Drive, Google Maps, PostgreSQL, Puppeteer, Redis, Sentry, Slack, SQLite | Do not install as live picks. Move to a maintained first-party or community replacement (below). |
The most important migration: the archived @modelcontextprotocol/server-github reference is replaced by GitHub's own official server, github/github-mcp-server (written in Go), which also offers a hosted remote endpoint. If a guide still tells you to install the old reference GitHub server, it is out of date.
What are the best MCP servers by category?
Here is the category-by-category shortlist, each pick chosen for maintainership and scope rather than novelty. Vendor capability descriptions below are vendor-reported unless otherwise noted.
Filesystem — official Filesystem server
The official Filesystem server (Node.js) is the right default for giving an assistant local file access. It performs secure file operations with configurable access controls, and you pass the directories it may touch either as CLI arguments or dynamically via MCP Roots. The non-negotiable: never give it your home directory — restrict it to explicit project folders. This is the single most common over-permissioning mistake in MCP setups.
Git and code hosting — official Git + GitHub's own server
For local repositories, the official Git server (Python, run via uvx) handles read, search and commit-level operations. For GitHub itself, use github/github-mcp-server: it ships toolsets for repositories, issues and pull requests, Actions CI/CD, and code/security analysis, and supports VS Code 1.101+, Claude Desktop, Cursor and Windsurf. Choose its remote hosted mode for zero local setup, or run it locally if you want your token to stay on your machine.
Web fetch and search — official Fetch (+ community search)
The official Fetch server retrieves and converts web pages to model-friendly text — the cleanest first-party option for “read this URL.” For actual search APIs, the old Brave reference server is archived, but the official repo now points to Brave's maintained replacement, brave/brave-search-mcp-server. For Perplexity, Firecrawl-style crawlers and other search wrappers, treat community servers as untrusted code until you have reviewed them.
Database — maintained Postgres-class server (not the archived one)
The official PostgreSQL and SQLite reference servers are archived — do not use them. If your app is on Supabase, use the Supabase MCP server with a project ref and read_only=true. If your app is on Neon, use Neon's first-party MCP server for branch and database workflows. For a generic Postgres database, use a maintained server only with a read-only database role and non-production data by default. A database server with write access plus prompt-injectable input is one of the highest-risk MCP configurations you can run; least-privilege here is not optional.
Browser automation — Playwright MCP
For driving a real browser (navigation, form fill, scraping rendered pages, end-to-end testing), use Microsoft/Playwright's maintained Playwright MCP. It is the browser automation pick now that the old Puppeteer reference server is archived. It is powerful and therefore high-blast-radius: a browser server that follows links can be steered by indirect prompt injection on a malicious page, so keep it scoped and supervised.
Memory and reasoning — official Memory + Sequential Thinking
The official Memory server gives an assistant a persistent knowledge graph across sessions — the canonical “remember facts about me/the project” primitive. Sequential Thinking exposes a structured, multi-step reasoning scratchpad. Both are local, low-risk, and a good first pair to install when you are learning how MCP wiring behaves before adding anything that touches the network or your files.
Remote (hosted) vs local servers — which transport?
Choose local stdio servers for runtime control, and remote (hosted) servers for zero-setup convenience. A local server runs as a child process on your machine, so the MCP process and token storage remain under your local environment — the trade-off is you manage the runtime and updates. A remote server, like GitHub's hosted endpoint, runs over HTTP with OAuth and needs nothing installed locally, but you are trusting the hosted service with whatever scope you grant. For sensitive repos or regulated data, default to local and least-privilege; for quick experimentation or shared team setups, remote is the lower-friction path. This local-vs-remote split is new in 2026 and barely mentioned in older roundups, yet it changes both your security posture and your setup time.
How do you add an MCP server to Claude, Cursor, or Claude Code?
For Claude Desktop you edit one JSON file; for Claude Code and Cursor you use a CLI/UI. The canonical Claude Desktop config lives at %APPDATA%\Claude\claude_desktop_config.json on Windows and ~/Library/Application Support/Claude/claude_desktop_config.json on macOS. Servers go under the mcpServers key, you must use absolute paths, and you must fully quit and restart Claude Desktop for changes to load. The commands below are taken from the official docs and are safe to copy.
That block is valid JSON, so it pastes straight into claude_desktop_config.json without edits — just swap in your real paths. Keep the filesystem argument list restricted to the explicit directories you want the assistant to touch (here, only /Users/you/Projects); never hand it your home directory or a drive root. Remember the OS-specific config location and the full quit-and-restart noted above.
TypeScript servers run via npx; Python servers run via uvx (or pip). For Claude Code and Cursor, you do not hand-edit JSON — you register the same servers from the command line, which is faster and less error-prone:
Which client should host these servers is a separate decision — if you are weighing the IDE/agent layer itself, see Cursor vs Claude Code and the broader best AI coding assistants for 2026 rundown. This page is about the servers; those pages are about the clients that connect to them.
How risky are MCP servers, and how do you harden them?
MCP servers are a real and active attack surface in 2026, so security is part of the pick, not an afterthought. OX Security reported 30+ MCP supply-chain disclosures and 10+ CVEs, and the named vulnerabilities are enough to justify conservative defaults: CVE-2025-6514 in the widely used mcp-remote proxy and CVE-2025-49596 in the MCP Inspector tool. Treat exact CVSS values and package-download counts as source-specific, and cross-check against the NVD or GitHub Advisory Database before quoting them in production docs.
The dominant attack classes are tool poisoning and indirect prompt injection. As Microsoft's engineering guidance explains, a malicious server (or malicious content a server fetches) can embed instructions in tool descriptions or returned data that the model then obeys — the model cannot reliably tell trusted developer instructions from attacker text injected through a tool. Security vendor OX Security has separately disclosed systemic remote-code-execution and supply-chain risk across MCP server packages. For EU teams, this also intersects governance: an MCP server with file, repo or database write access is part of your AI system's risk surface under the EU AI Act and existing security obligations, so document and scope it accordingly.
1) Prefer official or first-party servers over community wrappers. 2) Pin exact versions — do not float to latest. 3) Restrict the filesystem server to explicit allowed directories, never your home folder. 4) Give database servers read-only roles by default. 5) Review tool metadata and descriptions before enabling. 6) Do not run untrusted community servers with broad scopes or write access. 7) Be cautious with proxies like mcp-remote and dev tools like MCP Inspector — both have had critical RCE CVEs.
What does this mean in practice?
In practice, the winning MCP setup in 2026 is short, official-first, and tightly scoped — the opposite of the “install 20 servers” listicle. Start with Filesystem (locked to one project), Memory and Sequential Thinking to learn the wiring; add Git and GitHub's own server for development; add Playwright and a read-only Postgres-class server only when a task genuinely needs them. Skip every archived reference server. Where MCP fits in the larger picture — the agent loop that calls these tools — is covered in our AI agent architecture explainer, and if you are building the agent yourself, how to build an AI agent with the Claude Agent SDK shows the consumer side of these servers. For multi-agent setups that share tools, see multi-agent systems explained, and for the orchestration layer above the protocol, the best AI agent frameworks for 2026 and the LangGraph vs CrewAI vs AutoGen comparison both cover frameworks that integrate MCP tools.
Server status (7 live reference servers vs 14 archived legacy entries), GitHub's first-party server, config paths and install commands are verified against official Model Context Protocol, Claude Code and Cursor documentation. Security wording is grounded in OX Security research, Microsoft engineering guidance, NVD records and GitHub advisories; treat severity framing as source-reported and re-verify CVSS before quoting. The MCP Registry is in preview, not GA. Links accessed July 16, 2026.
FAQ
What are the official MCP servers in 2026?
There are seven current official reference servers: Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking and Time, maintained in the modelcontextprotocol/servers repository. Fourteen legacy reference entries — including GitHub, PostgreSQL, Slack, Brave Search, Google Drive and a legacy Git entry — were archived to a servers-archived repo and are historical reference only. The repo notes these are educational examples, not production-hardened products.
Is the official GitHub MCP server still maintained?
The old reference server (@modelcontextprotocol/server-github) was archived. GitHub now maintains its own official server at github/github-mcp-server, written in Go, with a hosted remote endpoint and local option. It supports VS Code 1.101+, Claude Desktop, Cursor and Windsurf, and ships toolsets for repositories, issues and pull requests, Actions CI/CD, and code/security analysis. That is the correct GitHub MCP pick today.
Where is the Claude Desktop MCP config file?
On Windows it is %APPDATA%\Claude\claude_desktop_config.json; on macOS it is ~/Library/Application Support/Claude/claude_desktop_config.json. Add servers under the mcpServers key, use absolute paths, and fully quit and restart Claude Desktop for changes to take effect. For Claude Code and Cursor you register servers via the CLI (claude mcp add) or the app’s MCP settings instead of editing JSON by hand.
Are MCP servers safe to run?
Only with care. OX Security reported 30+ MCP supply-chain disclosures and 10+ CVEs, and named criticals include CVE-2025-6514 in mcp-remote and CVE-2025-49596 in MCP Inspector. Tool poisoning and indirect prompt injection are the main attack classes. Prefer official servers, pin versions, restrict scopes, and avoid untrusted community servers with broad permissions.
What is the MCP Registry?
The official MCP Registry, at registry.modelcontextprotocol.io, is the canonical discovery index for MCP servers — described at launch as an app store for MCP servers. It went live in preview on 8 September 2025 and is still preview, with no data-durability guarantee and general availability pending. Use it to find servers, but it is an index, not a security guarantee — still vet each server yourself.
How do I add an MCP server to Cursor?
In Cursor, open Settings, go to the MCP section, and add a server, or edit ~/.cursor/mcp.json directly. Cursor uses the same {“mcpServers”: {…}} JSON shape as Claude Desktop, so a config that works in one can usually be reused in the other. TypeScript servers run via npx; Python servers via uvx or pip. Restart Cursor after adding a server.
Which MCP servers should a beginner install first?
Start with low-risk local primitives: the official Filesystem server locked to a single project directory, plus Memory and Sequential Thinking. These teach you how MCP wiring and tool calls behave without exposing the network or your whole machine. Add Git and GitHub’s first-party server for development, then Playwright (browser) and a read-only Postgres-class server only when a specific task requires them.
Bibliography (16 sources)
Sources prioritise official Model Context Protocol documentation, first-party vendor documentation and public vulnerability databases. Security statistics and CVE scores are source-reported; re-verify CVSS against NVD or GitHub Advisory Database before quoting. The MCP Registry is in preview, not GA. Links accessed July 16, 2026.
- Model Context Protocol — Example Servers (official docs). Authoritative list of the 7 current reference servers and canonical Claude config JSON. modelcontextprotocol.io/examples
- Model Context Protocol — modelcontextprotocol/servers (GitHub). Confirms 7 active reference servers and 14 archived legacy entries; states they are educational examples, not production-ready. github.com/modelcontextprotocol/servers
- GitHub — github/github-mcp-server (official GitHub MCP server, Go). First-party server with remote option; toolsets for repos, issues/PRs, Actions, code/security analysis. github.com/github/github-mcp-server
- Model Context Protocol — Introducing the MCP Registry (blog, 8 Sep 2025). Registry launched in preview as an “app store for MCP servers.” blog.modelcontextprotocol.io
- Model Context Protocol — Official MCP Registry. Canonical live discovery index for MCP servers (preview). registry.modelcontextprotocol.io
- Model Context Protocol — Filesystem server README. Node.js secure file ops with configurable access controls; directories via CLI args or Roots. github.com/modelcontextprotocol/servers (filesystem)
- Model Context Protocol — Connect to local MCP servers. Current local-server configuration flow and client setup concepts. modelcontextprotocol.io/docs/develop/connect-local-servers
- Anthropic — Connect to MCP servers (Claude Code docs). The
claude mcp addflow for local and remote servers. code.claude.com - Cursor — MCP docs. Cursor's Settings / Tools & MCP and
mcp.jsonsetup. docs.cursor.com/en/tools/mcp - Microsoft Playwright — Playwright MCP. First-party browser automation server and installation docs. playwright.dev/docs/getting-started-mcp
- Supabase — Supabase MCP server. Vendor MCP server with project scoping and read-only mode guidance. supabase.com/docs/guides/ai-tools/mcp
- Neon — Neon MCP Server. First-party MCP server for Neon database and branch workflows. neon.com/docs/ai/neon-mcp-server
- OX Security — MCP Supply Chain Advisory: RCE Vulnerabilities Across the AI Ecosystem. Research disclosure of systemic RCE/supply-chain risk and 30+ disclosures / 10+ CVEs framing. ox.security
- Microsoft for Developers — Protecting against indirect prompt injection in MCP. Engineering guidance on tool poisoning / indirect injection and mitigations. developer.microsoft.com
- NVD — CVE-2025-6514. Vulnerability record for
mcp-remote. nvd.nist.gov - NVD — CVE-2025-49596. Vulnerability record for MCP Inspector. nvd.nist.gov
