Best MCP servers and Claude tools to install first

The short answer: install Anthropic’s reference servers first, add the GitHub server if you write code, and add Playwright if you want Claude driving a browser. After that, you add by job. Running issues and docs? Linear and Notion. Running a real business process? That’s where a server like Tallyfy comes in.

That covers most people. Three or four servers, well chosen, beats a sprawling pile you forget you installed.

But here’s the part the “top 50 MCP servers” lists skip. A model with no servers connected can only talk back at you. MCP is the standard, basically, that lets Claude reach into your files, your repos, and your tools and do something useful with them. The list of servers matters far less than picking the handful you’ll actually open every day.

What makes a server worth installing

One disclosure up front, because it shapes how you should read this. We run Tallyfy, which is workflow software, and we operate an MCP server in production, so I’ve spent real time wiring these servers into Claude myself. That gives me a bias and also some scar tissue. Like the rest of our software tool breakdowns, I’ll tell you plainly where each server is worth a spot on your machine and where it just adds clutter you’ll never open.

The criteria here are deliberately dull. Who maintains it, official or community? What does it actually let Claude do? Does it run locally on your machine or as a remote hosted endpoint? And does it need an API key or an OAuth login to work? Those four questions sort the genuinely useful servers from the demos.

What nobody tells you until you’ve installed your fifth or sixth server is that the cost isn’t the install. It’s the upkeep. Each local server is another process to keep alive, another set of credentials, another thing that breaks quietly when a version bumps. So the honest goal is fewer servers doing more.

A trophy wall of integrations impresses you once, then quietly rots.

Here’s the whole shortlist at a glance, grouped the way you should think about it.

Start with the servers Anthropic maintains

If you install nothing else, install these. The official reference servers are maintained by Anthropic, built with the community, and they cover the boring fundamentals every other workflow leans on. The repo describes them in one line each, and that plain language is the point: these are tools, not platforms. They run locally as small processes, you point them at a folder or a repo, and Claude can suddenly read and act on what’s actually on your disk instead of guessing from what you paste into chat.

The five to know are Filesystem (“Secure file operations with configurable access controls”), Git (“Tools to read, search, and manipulate Git repositories”), Fetch (“Web content fetching and conversion for efficient LLM usage”), Memory (“Knowledge graph-based persistent memory system”), and Sequential Thinking (“Dynamic and reflective problem-solving through thought sequences”). Two more, Time and Everything, ship alongside them, but those five are the daily drivers.

Filesystem is the one you install on day one. It’s the difference between Claude reasoning about your project and Claude actually reading it. Git pairs naturally with it for anyone working in a repo. Fetch turns out to be quietly useful for pulling a doc page or a changelog straight into the conversation, instead of you copy-pasting it across.

Boring, all five of them, and the whole stack leans on them anyway.

Memory and Sequential Thinking are the two people skip and then come back for. Memory gives a long session a knowledge graph it can write to, so context survives past a single chat. Sequential Thinking nudges the model to break a gnarly problem into ordered steps. Neither is flashy. Both quietly raise the floor on harder work.

Which third-party servers come next?

Once the basics are in, the next tier is where MCP starts to feel less like a toy. These are maintained by the companies whose products they connect to, which matters more than it sounds: an official server tends to track the real API and survive the next breaking change, where a one-off wrapper that someone had to cobble together in a weekend usually doesn’t. The four below are the ones I would reach for before any of the thousands of community servers floating around. Each one turns a system you already use into something Claude can read and change directly, which is the entire promise of the protocol. They split cleanly by what you do all day, so install the ones that match your work and leave the rest. A server you never call is just one more thing to patch when it breaks.

GitHub MCP

The GitHub MCP server is GitHub’s own, and for anyone who writes code it’s the highest-impact install after the filesystem server. It “connects AI tools directly to GitHub’s platform,” which in practice means Claude can read repositories, manage issues and pull requests, review code, and watch Actions runs, all without you leaving the conversation. It comes in two flavors: a remote version GitHub hosts for you, which is the easiest setup, and a local Docker build for locked-down environments. With roughly thirty thousand stars, it’s also one of the most battle-tested servers in the ecosystem.

Playwright MCP

Playwright MCP is Microsoft’s browser-automation server, and it made one design choice that makes it stand out: “Uses Playwright’s accessibility tree, not pixel-based input.” No screenshots, no vision model squinting at pixels. It drives the page off structured accessibility data, which is faster, cheaper on tokens, and far less janky than the screenshot-and-click approach. If you want Claude filling forms, scraping a page, or testing a flow, this is the one. It carries well over thirty thousand stars for a reason.

Cloudflare and Brave Search

The other two are narrower but worth knowing. Cloudflare ships thirteen separate hosted servers across its products, covering Workers builds, DNS analytics, observability logs, browser rendering, and more, each at its own remote endpoint, so you connect only the slice you need. Brave Search gives Claude live web, news, image, and video search through Brave’s API, which is the clean way to get current information into a session instead of relying on stale training data. It needs a free API key to run, the one bit of friction in an otherwise quick setup.

Past those, if your work means querying data, there are solid community servers for talking to a Postgres or SQLite database directly, so Claude can answer a real question against your tables instead of you exporting a CSV and pasting it in. Just check who maintains the one you pick before you hand it a connection string.

Adding a remote server takes one line. Here’s the actual syntax from the Claude Code docs, for a hosted server and a local one:

# A remote, hosted server (the recommended transport)
claude mcp add --transport http notion https://mcp.notion.com/mcp

# A local server that runs as a subprocess on your machine
claude mcp add --transport stdio airtable --env AIRTABLE_API_KEY=YOUR_KEY -- npx -y airtable-mcp-server

Run claude mcp list afterward and the server shows up with its tools ready to call. That’s the whole ceremony.

When the work isn’t writing code

Most MCP roundups stop at developer tools, as if the only people connecting Claude to things are engineers. That misses the bigger shift. The same protocol that hands a coding agent your repo can hand an operations team their actual systems of work, and the official servers for that side of the house have quietly arrived. Slack, Linear, and Notion all now run their own hosted servers, so Claude can search a workspace, file an issue, or update a doc with admin-approved access instead of a brittle bot token someone set up two years ago. This is the part that turns Claude from a coding sidekick into something an ops, support, or marketing team can lean on. The trick is the same as before: connect the one or two systems your team genuinely lives in and skip every server with a logo.

An ops lead with Linear and Notion wired in gets the same reach engineers got first.

Slack’s MCP server is generally available and lets assistants like Claude search messages, find information, and take actions, with workspace admins approving each integration. Linear’s server at its hosted endpoint covers finding, creating, and updating issues, projects, and comments, which is most of what an engineering-adjacent team asks of it. Notion’s hosted server gives Claude “secure access to your Notion workspace,” reading and writing pages the way you would yourself. There’s also a Zapier MCP server that fans out to thousands of apps through Zapier’s connector library, handy as a catch-all, though it inherits the brittleness of any connector-era middleware.

None of these needed an engineer to switch on.

A misconception we keep running into about MCP is that more servers means more capability. It usually means the opposite. The teams getting real value connect two or three systems and go deep, while the ones with twenty servers spend their time debugging the mesh. Pick the systems where your work actually happens.

Tallyfy MCP

Honorable mention here, and I’ll be upfront that this is ours, so weigh it accordingly. Most of the servers above connect Claude to a tool: a repo, a doc, a channel. The Tallyfy MCP server connects Claude to a running process instead. It exposes 100+ tools across the platform, uses OAuth 2.1 for auth rather than a pasted token, and it’s listed in the Official MCP Registry as com.tallyfy/mcp-server. The point isn’t the tool count. It’s the shape. A coding server lets an agent change a file; a process server lets an agent move a real piece of work forward, with a defined next step and a human owner where one is needed.

That distinction matters more as agents take on real tasks. Connecting Claude to your tools is the start. Giving those tool calls a process to run inside, with automations and conditional steps deciding what happens next, is what keeps an agent from wandering off. If you want the deeper version of this argument, our guide to the Tallyfy MCP server walks through it, and mcp-server-sprawl makes the case for keeping the count low.

Pick the editor or host that runs Claude

A server is only half the setup. Something has to host Claude and call those servers, and your choice of host shapes the whole experience. The host decides how you add a server, whether that config travels with your team or stays stuck on your laptop, and how much friction sits between you and a working tool.

The natural starting point is Claude Code, Anthropic’s official CLI, which speaks MCP directly through the claude mcp add command you saw above and is built for exactly this. But you’re not locked into it. A handful of editors and agents now run Claude and connect MCP servers too, and which one fits depends on whether you live in a terminal, an editor, or a desktop app. None of these is wrong. They just suit different habits, and most are free or open source to try.

For an editor-first workflow, Cline is the open-source coding agent for VS Code, with more than eight million installs and built-in MCP support, all under an Apache 2.0 license. Cursor is the popular AI editor that runs the latest Claude models and bills itself as “your coding agent for building ambitious software.” Zed, the fast Rust-based editor, leans into “Agentic Editing” and lets you connect MCP servers to extend what the agent knows. If you prefer the terminal, Aider is “AI pair programming in your terminal” and is explicitly tuned to work best with Claude.

Try two of these for a week and keep the one you stop noticing you’re using.

There’s also a gentler on-ramp for people who would rather not touch a config file. Claude Desktop now installs MCP servers as one-click bundles. Anthropic’s Desktop Extensions bundle “an entire MCP server … into a single installable package,” dependencies and all: you download a .mcpb file, double-click, and click install. No terminal, no JSON. For a non-developer who just wants Claude talking to a tool, that’s about as easy as setup gets, and it’s where a lot of business users will start.

So the real first move isn’t picking a server. It’s picking the host. Get Claude Code or Claude Desktop running, wire in the filesystem server, and add the next one only when you hit a wall that a server would solve. That order keeps the setup honest and the mess down.

FAQ