There are now thousands of MCP servers listed across the various directories, and most of them are either abandoned side projects or thin wrappers someone shipped in an afternoon in 2025 and never touched again. This isn’t a directory dump. It’s the list we’d actually hand a colleague who asked “what should I connect to Claude/Cursor this week” — one honest paragraph per pick, official-vs-community status called out, and a section on what we deliberately left off and why.

We build and run our own MCP server, so we’ve spent more time than average reading other people’s OAuth flows, tool schemas, and changelogs. That bias shows up in what we picked: things that are maintained as of mid-2026, official where an official option exists, and useful enough that we or someone we trust actually reaches for them daily.

How we picked the best MCP servers

Three criteria, in order:

  1. Maintained, as of mid-2026. A recent release or commit, not just a GitHub star count from 2024. Anything with an “archived” banner or a deprecation notice pointing elsewhere is excluded from the main list (see “what we dropped” below).
  2. Official where possible. If the vendor ships a first-party MCP server, we default to recommending that over a community wrapper around the same API — official servers get the vendor’s own auth, rate limits, and support commitment behind them. Community picks are included where they fill a real gap or where there’s no first-party alternative.
  3. Actually useful daily, not “technically works in a demo.” We favor servers with tool surfaces broad enough to replace a browser tab, not single-tool novelty wrappers.

Every status claim below (official/community, active/deprecated, star counts, release dates) is checked against the project’s own repo or vendor docs, not secondhand blog posts, as of July 2026. Things move fast in this ecosystem — recheck before you build critical infra on any of these.

Dev & code

GitHub MCP server — Official, maintained by GitHub itself, at v1.7.0 as of late July 2026, with consolidated Projects tools and OAuth scope filtering shipped earlier this year. Lets an agent read repos and code, manage issues and PRs, watch Actions runs, and review security alerts directly against GitHub’s API. This is the obvious first install if you do any coding work inside Claude Code, Cursor, or a similar client — there’s no meaningful community alternative worth using over the first-party one.

Sentry MCP server — Official, hosted remote server at mcp.sentry.dev/mcp with OAuth, launched March 2025 and steadily expanded since. Exposes about 20 tools across error, performance, and release data — deliberately few: the repo’s own agent guidelines cap the server at 25 tools so it doesn’t eat an agent’s tool slots, even though the codebase defines roughly twice that many under the hood. An agent can pull the latest unhandled exception, read the actual stack trace, and propose a fix without you copy-pasting a Sentry link into the chat. Good fit if Sentry is already your error tracker; there’s no reason to route around it.

Linear MCP server — Official, remote and OAuth 2.1-authenticated at mcp.linear.app/mcp, built by Linear with Cloudflare and Anthropic, with 25+ tools for issues, projects, and comments. It’s been generally available since mid-2025, not new this year, but it’s stable and does what you’d expect: an agent can read a ticket, update its status, and leave a comment without you leaving your editor.

Serena — Community, MIT-licensed, ~27k GitHub stars, last release in late July 2026. A Python-based coding-agent toolkit that adds IDE-like semantic code retrieval and editing at the symbol level — find a function, rename it, refactor its call sites — instead of dumping whole files into context. If your agent’s coding workflow feels token-hungry and imprecise on a large codebase, this is the community project worth trying before you write your own.

Docs & context

Context7 — Official, from Upstash. Injects up-to-date, version-specific library documentation and code examples into context via a resolve-library-id plus get-library-docs tool pair, aimed squarely at the “the model hallucinated an API that was deprecated two versions ago” problem. Available as a remote endpoint (mcp.context7.com/mcp) or a local @upstash/context7-mcp npm package. Cheap to add, genuinely reduces stale-API hallucination for fast-moving libraries.

Reference filesystem and memory servers — Official, from the MCP project itself (modelcontextprotocol/servers). Worth a specific callout because most of the original first-party reference servers were archived and handed off to vendors or the community during 2025 — filesystem and memory are among the small set that survived as maintained upstream references, both shipping regular releases as of July 2026. Filesystem gives an agent sandboxed local read/write with configurable access controls; memory is a minimal knowledge-graph-style persistent memory implementation, explicitly documented as a reference, not a production memory store. Good starting points if you’re building your own server and want to see how the MCP maintainers structure a minimal one.

Browser & automation

Playwright MCP — Official, from Microsoft, ~35.5k GitHub stars, last release in early July 2026, Apache-2.0, and listed in the official MCP Registry. Browser automation via the Playwright engine exposed as MCP tools: navigate, click, fill forms, read the rendered DOM. This is the current default for “let an agent drive a real browser.” Microsoft also shipped a companion @playwright/cli in early 2026 as a lower-token-cost non-MCP alternative for the same automation, worth knowing about if MCP’s tool-call overhead (roughly 4x more tokens per action, by Microsoft’s own framing) matters for a high-volume automation job.

Zapier MCP — Official, hosted remote server at mcp.zapier.com (Streamable HTTP under /api/mcp/v1), exposing 9,000+ connected apps and 30,000+ actions (per Zapier’s own MCP landing page) as MCP tools through a single OAuth connection. If you need “send this to Slack, then update this spreadsheet, then create this calendar event” without writing a per-app integration, this is the fastest path.

n8n’s native MCP server — Official, in Public Preview since April 2026 and bundled in every n8n edition (Cloud, Enterprise, self-hosted Community). It lets an agent build, test, and publish n8n workflows via MCP. Worth being precise here because there’s a second, unrelated thing with a similar name: the community npm package n8n-nodes-mcp lets n8n workflows call out to external MCP servers — that’s n8n acting as an MCP client, the opposite direction from the native server. If you want an agent to build automations in n8n, you want the native server; if you want your n8n workflows to call other people’s MCP tools, that’s the community node.

Data & analytics

Snowflake’s managed MCP server — Official, runs inside Snowflake with OAuth, exposing Cortex Analyst, Cortex Search, Cortex Agents, and custom SQL as MCP tools with no separate infrastructure to run. Note: the older Snowflake-Labs/mcp open-source project is now deprecated in favor of this first-party server — if you find blog posts pointing at the Labs repo, they’re describing the predecessor, not the current recommended path.

PostHog MCP — Official, with 40+ tools (per the repo’s tool definitions as of our check) spanning analytics queries, feature flags, insights, experiments, error tracking, and SQL, hosted at mcp.posthog.com. Also ships as a Claude Code plugin with slash commands. If you’re already on PostHog for product analytics and feature flags, this is a clean way to query it conversationally instead of building dashboards for one-off questions.

mcp-analytics — This one is ours, so take the placement with appropriate skepticism: we’re listing it because it fits the category, not because it’s the best pick in it. mcp-analytics is web analytics whose only interface is an MCP server (23 tools, OAuth 2.1 with PKCE, no dashboard by design) — you ask Claude, ChatGPT, or Cursor things like “how much of last week’s traffic was real humans versus AI crawlers” and get an answer via tools like traffic_class_breakdown and top_pages. It’s narrower than PostHog or Snowflake — site traffic and bot classification, not product analytics or a general data warehouse — so if you need funnels or feature flags, PostHog is the better fit; if you need warehouse-scale SQL, Snowflake is. If you specifically want to know what’s crawling your site and whether it’s a real visitor, an AI agent browsing for a human, or a training crawler, that’s the gap we built this for. Full tool catalog at /mcp/tools.

Fun / desktop

Blender MCP — Community, independent (built by Siddharth Ahuja), ~24.8k GitHub stars, with open issues as recent as July 2026. Connects Blender to an LLM — originally built for Claude — so the model can directly drive Blender for prompt-assisted 3D modeling and scene creation. Not something most people need daily, but a clear demonstration of MCP doing something a REST API wouldn’t make pleasant: turning a chat message into live tool calls against a stateful 3D application.

Honorable mention: Cloudflare’s OAuth reference

Cloudflare’s MCP servers and workers-oauth-provider — Official, from Cloudflare. Beyond their own ready-made servers (a Workers Bindings server with ~25 tools for managing Workers infrastructure), Cloudflare ships workers-oauth-provider, a library any Worker-hosted MCP server can use to add OAuth. We didn’t build on Workers ourselves, but we benchmarked our own OAuth implementation’s design choices against theirs repeatedly while getting our connector working — see our OAuth deep dive if you’re implementing this yourself and want to know exactly which spec-permissive choices to copy.

What we dropped, and why

Puppeteer MCP server. Was the original official reference automation server from the MCP project. As of 2026 it’s archived — moved to modelcontextprotocol/servers-archived, and the @modelcontextprotocol/server-puppeteer npm package itself says “no longer supported.” Playwright MCP is the maintained replacement and does the same job with an actively developed engine underneath. If a guide or a directory still points you at Puppeteer MCP, that guide is stale.

The original Snowflake-Labs/mcp repo. Deprecated in favor of Snowflake’s own managed MCP server, covered above. The community repo still exists on GitHub and still gets found by search, which is exactly the trap: it’s not maintained as the recommended path anymore, even though nothing about the repo visually screams “don’t use this.”

Everything we couldn’t verify was still maintained. Several servers on other “best of” lists from 2025 didn’t make it here because we couldn’t confirm a 2026 release, commit, or vendor support statement. Being missing from this list isn’t a verdict that a server is bad — it’s a signal we couldn’t confirm its maintenance status as of July 2026. Recheck before you depend on anything, including our own picks above; this is a fast-moving list by nature.

How to try any of these

If you haven’t connected an MCP server to Claude before, start with our Claude MCP setup guide — it walks through adding a custom connector end to end. If you’re building your own server rather than just using someone else’s, MCP Inspector is the tool for poking at a server’s tools and schemas before you wire it into a real client, and if you’re deciding whether MCP is even the right layer for what you’re building versus a plain API, we wrote a comparison covering what MCP adds and what it doesn’t replace.