Best MCP servers in 2026: a founder's shortlist

The Model Context Protocol turned eighteen months old in May 2026. There are now more than two thousand MCP servers on GitHub. Most are abandoned. Many overlap. None come with the trust signals you would expect from a normal package registry. Choosing one is currently an afternoon of testing.

This is the shortlist we actually use across Claude Code, Cursor and Claude Desktop. The selection criteria were boring on purpose: maintained in the last 60 days, install success rate above 90% on macOS and Linux, clear documentation of the tools it exposes, and at least one production user we could verify.

How we picked

Three signals matter more than star counts on GitHub:

• Last meaningful commit. Not a README typo. A real fix or feature in the last 60 days.
• Issue triage. Open issues that get a response within seven days. Dead trackers are an instant fail.
• Tool surface area. Servers that expose 3-15 well-scoped tools beat servers that expose one giant tool or fifty trivial ones.

We did not weight star count at all. Many of the highest-starred servers are demos. Several of the most-used servers below have under 500 stars.

The eleven we keep installed

1. filesystem (Anthropic-maintained)

The reference server. Read, write, search local files with scoped permissions. Use it as the baseline for everything else.

2. github (Anthropic-maintained)

Issue, PR and search APIs. The only server we trust for production GitHub work because the auth flow is correct and the rate-limit handling is sane.

3. postgres (community, well-maintained)

Read-only by default, which is the right default. Switching to write needs an explicit flag. Saves you from "let me just inspect this table" turning into a destructive query.

4. puppeteer / playwright

For deterministic browser automation. Pick playwright if you are on a modern stack, puppeteer if you need legacy Chrome quirks. Both have working MCP wrappers; both ship with screenshot tools that matter for verification.

5. linear

The official Linear MCP is now stable. Replace it only if you have specific webhook needs the official one does not support.

6. notion

The official Notion MCP. Earlier community wrappers tripped on databases with rollups and formulas. The official one handles them.

7. slack

Read messages, post messages, manage scheduled sends. Be careful with the post tool: scope it to a single channel during testing.

8. brave-search or perplexity

Either one. Both expose web search to your agent. Brave is cheaper, Perplexity is better at synthesised answers. We use both for different tasks.

9. supabase

The official Supabase MCP. Schema, migrations, RLS policy inspection. The killer feature is being able to read your live schema before you write a migration.

10. obsidian or apple-notes

For personal knowledge agents. Obsidian if you live in markdown, Apple Notes if you live on iOS. Both are mature enough to trust with daily writing.

11. memory (Anthropic-maintained reference)

A simple key-value memory layer. Not the right answer for production but the right answer for one-person workflows where you want continuity across sessions.

What we removed in the last quarter

Equally important: three servers we previously recommended and have since uninstalled.

• Generic "scraper" MCPs. Replaced by a hardened playwright setup, which is more reliable and easier to audit.
• Most "all-in-one" servers. Servers that bundle ten unrelated tools fail one of them and then the LLM gets confused about which tool to call. Single-purpose servers win.
• Servers requiring a SaaS account just to use locally. If a server demands a hosted account before it will start, the friction is not worth it.

The bigger pattern

The MCP registry problem is so painful we listed it as AI14 in our directory. A reviewed registry with actual install-success signals, a per-server compatibility matrix and an AI-assisted "paste your task, get the right MCP" recommender would be a clear win.

Until someone builds it, the shortlist above is the most useful filter we know.