The best MCP servers in 2026

Search "best MCP servers" and you get a GitHub list with 1,400 entries, no opinion, and no way to tell which ones are maintained. This is a different kind of MCP server list: we track 57 tools shipping an official Model Context Protocol server in our directory, we maintain a graveyard of 219 AI tools that died, and we run an MCP server ourselves. Below are the seventeen worth connecting first in 2026 — one per job, with what each actually gives your assistant and where it disappoints. One of them is ours, disclosed and ranked last.

What is an MCP server?

The Model Context Protocol is an open standard for connecting AI assistants to outside tools and data. An MCP server is the small adapter a company publishes so an assistant like Claude can call its product directly — read your repository, query your database, charge a card — instead of you copying context in and results out by hand.

The distinction that matters when picking: an MCP server gives a model actions and live data, not just knowledge. That is why the install decision is a permissions decision. Every server below is judged on what it can reach, not only on what it can do.

Start here: GitHub

The GitHub server is the one to install first, and for many developers it is the only one they need for a month. It gives an assistant your repositories, issues, pull requests, and Actions runs — which turns "here's my error, let me paste the file" into "look at the failing run and open a PR." Scope the token narrowly; the default install is more access than most sessions need.

Production errors: Sentry

Sentry's server closes the loop the GitHub one opens. The assistant pulls the actual stack trace, the release it appeared in, and how many users hit it, then goes and fixes the code. Debugging from real production data instead of a pasted snippet is the single biggest quality jump most teams get from MCP.

Product analytics: PostHog

PostHog lets an assistant query your product data conversationally — funnels, retention, feature-flag state, session counts — without you building the dashboard first. It is the best example of MCP's underrated use case: not writing code, but answering questions that used to require someone who knows the query language.

Databases: Neon and Prisma

Neon exposes serverless Postgres — branches, queries, migrations — so an assistant can inspect a schema and reason about real rows rather than guessing at your data model. Prisma covers the ORM layer, which matters if your schema lives in Prisma's own format. Between them, most modern TypeScript stacks are covered. Read-only credentials for anything pointed at production, always.

Payments: Stripe

Stripe's server is the clearest example of MCP moving from reading to doing: create a customer, generate a payment link, look up a subscription, all from the assistant. It is also the clearest example of why blast radius matters. Use restricted API keys, and test in sandbox mode before you ever point it at live.

Deploys and edge: Cloudflare

Cloudflare publishes a set of servers covering Workers, DNS, and analytics — the deploy-and-diagnose half of a web stack. Paired with GitHub and Sentry it makes the whole loop legible to an assistant: what shipped, what broke, and what the edge saw.

Docs and workspace: Notion

Notion AI's server is the highest-value non-engineering install on this list. It lets an assistant read and write your team's actual docs, specs, and wikis, which is the difference between generic output and output that matches how your company already writes things down.

Issue tracking: Linear

Linear gives an assistant your issues, projects, and cycles — so "what's blocking the release" becomes a question you ask rather than a board you read. It is the tightest, most predictable server on this list; Linear's API discipline shows.

Everything else: Zapier

Zapier's MCP server is the pragmatic answer for the 8,000 apps that will never ship a first-party server. It bridges your assistant to whatever your business already runs — CRM, email, spreadsheets, calendars — at the cost of an extra hop and Zapier's task pricing. When a vendor has no official server, check here before concluding it can't be done.

Browser automation: Browserbase

Browserbase runs headless browsers in the cloud with session recording, which is what agents need when a task requires clicking through a real site rather than calling an API. It is the most reliable of the browser-infrastructure options we track, and its Stagehand framework covers the harder scripted cases.

Reading the web: Firecrawl

Firecrawl turns any URL into clean markdown an assistant can actually use — handling JavaScript rendering, pagination, and the mess real sites are made of. If your workflow involves "read these forty pages and summarize," this is the server that makes it work instead of half-work.

Search built for agents: Exa

Exa is a search API designed for models rather than humans — semantic queries, full page contents returned inline, no scraping step. General assistants already search; Exa is what you add when the searching needs to be systematic, filtered, and repeatable inside a workflow.

Current library docs: Context7

Context7 fixes the most common AI coding failure that isn't the model's fault: training-data staleness. It pulls the current documentation for the exact library version you're using, straight into context, which kills the class of bug where an assistant confidently writes an API call that was removed two releases ago.

Sandboxed code execution: E2B

E2B gives an assistant a real, isolated sandbox to run code in — data analysis, script testing, anything you don't want executing on your machine. It is the correct answer to "can the model just run this?" and, given this year's news about agents escaping their sandboxes, the correct place to run untrusted work.

Memory across sessions: Mem0

Mem0 adds persistent memory, so an assistant remembers preferences, project context, and prior decisions between conversations. Memory is the least mature category on this list — evaluate it on whether your recall improves, not on benchmark claims.

Our own: the ToolDirectory.AI MCP server

Disclosure, because it belongs at the bottom of a list we wrote: we publish a free, read-only MCP server that lets an assistant search our catalog of 2,600+ AI tools, pull structured details, and query the graveyard. It is useful for exactly one job — answering "what tool does X, and is it still alive" without leaving your assistant. It ranks last here because we built it.

The "awesome MCP servers" problem

The community awesome-MCP-servers lists are genuinely useful as raw inventory, and genuinely bad as a shortlist. They index everything ever published, including servers whose last commit was fourteen months ago and servers built by one person over a weekend against an API that has since changed. Nothing marks the dead ones.

That is the gap this list exists to fill. Everything above is a server published or officially supported by the company whose product it exposes, from a tool that is active in our directory today. When a listed tool dies, it moves to our AI graveyard and comes off this page — the same discipline we apply to every other list we publish.

How to add an MCP server to Claude

Adding MCP servers to Claude takes one command per server in Claude Code (claude mcp add), or the Connectors settings pane in the Claude desktop and web apps, where official servers can be enabled with an OAuth login. Claude Desktop also supports a local config file for servers you run yourself. Other MCP clients — Cursor, Windsurf, VS Code's agent mode — use the same protocol, so a server that works in one generally works in the others.

Two rules worth adopting before you install a dozen of them:

1. Least privilege, every time. Read-only tokens by default; scoped keys for anything that writes; sandbox keys for anything that moves money.
2. Fewer, better servers. Every connected server adds tool definitions the model must weigh on every request. Six well-chosen servers beat twenty, both for cost and for the assistant's decision quality.

How we picked

Every pick is active in our directory, carries a hand-maintained editorial review, and publishes an official or company-supported MCP server — no community forks of first-party APIs. We ranked by job coverage rather than popularity: one clear winner per job an assistant is actually asked to do, with the trade-offs named. No placement on this list is paid. The full field, including the servers that didn't make the shortlist, is the MCP servers category — 57 tools and growing. Our state of MCP servers report covers the adoption data behind it.