The MCP servers actually worth connecting to your assistant

Filed under Guide

By Gerald · 2 August 2026

Close-up of an audio mixing console with rows of faders, representing the process of dialing in which connections actually matter

I have connected close to thirty MCP servers to one assistant or another over the past year. I am actively using four of them. The rest were either redundant with what the model already does well, or they added a permission prompt to a task that used to take one sentence.

Most "best MCP servers" lists are just the registry sorted alphabetically with a one-line description copied from each server's README. That is not a recommendation, it is an index. This is a recommendation, based on servers I have actually run, in the clients I actually ran them in.

Connect a server when it changes a decision you make. Do not connect one just because it exists.

How I judged these

The Model Context Protocol got a real registry in late 2025, hosted at registry.modelcontextprotocol.io, and by mid-2026 the maintainers had already gone through one significant cleanup. The official servers repository now actively maintains only seven reference servers: Everything, Fetch, Filesystem, Git, Memory, Sequential Thinking, and Time. Twelve others, including GitHub, GitLab, Slack, Postgres, Sentry, and Google Drive, were archived and moved to a separate repository, replaced by versions maintained by the vendors themselves or by third parties. Brave Search now has an official server from Brave. Slack's is maintained by Zencoderai. That is worth knowing before you install anything: the "official" servers list quietly got smaller, and vendor-maintained versions are now the safer default for the categories that still have one.

My test for whether a server earns a permanent connection is simple. Does having it connected change a decision I make, or does it just move data from one place to another slightly faster than I could have done it myself. A weather server that reads today's forecast is not doing anything the model cannot already tell you to check. A server that reads your actual open tasks and tells you what is overdue is doing something the model genuinely could not know without it.

I also only recommend servers I have tested directly, and I say which client I tested them in, because MCP support is not identical across Claude Desktop, Claude Code, ChatGPT's developer mode, and Cursor. A server that installs cleanly in one sometimes needs an extra config step in another.

Notes and tasks: the connector that changed the most

Pliers and wrenches hanging on a workshop wall with their outlines marked behind them — The useful MCP servers are the ones with an obvious place in your workflow, not the ones that just look impressive on a list.

This is the one I actually use daily. I connected Flow's own MCP server, which exposes my notes, tasks, and captures to whatever assistant I am talking to, tested mainly in Claude Desktop and Claude Code.

The change was not speed. Copy-pasting one task into a chat is still faster than a tool call. The change was that the assistant stopped guessing. Ask a model with no context what you should work on this week and it gives you generic advice. Ask one that can read your actual board, your actual overdue items, and the note you wrote after last week's review, and it gives you an answer specific to your situation. If you want the full mechanics of that setup, I wrote a separate guide to connecting an assistant to your notes and tasks, and a more skeptical look at why use MCP at all when copy-paste is often faster.

Any tool that holds your live work state benefits from this same pattern: a task manager, a CRM, a project tracker. The category is "give the model a current picture of what you are actually doing," not any one product.

Files and local search

Filesystem is one of the seven servers Anthropic still maintains directly, and it is the single most useful one if you write code or manage documents locally. It lets the assistant read, search, and sometimes write files in a directory you approve, instead of you pasting file contents into the chat one at a time.

I run this constantly in Claude Code, less often in Claude Desktop where I am usually working from notes rather than a folder of source files. The productivity gain is real and it is boring: fewer copy-paste round trips, faster multi-file edits, search across a project without opening every file yourself.

Git is the other maintained reference server worth having if you work in a repository day to day. It lets the assistant check status, diffs, and history without you narrating the state of your branch every time you ask a question.

Web fetch and search servers

Fetch, also on Anthropic's actively maintained list, retrieves and reads a URL you give it. It is unglamorous and it is the one that quietly prevents the most wrong answers, because it lets the model check a page instead of recalling a stale version of it from training.

For actual search rather than a single URL, a dedicated search server matters more than people expect. Brave's own server, or a wrapper around a search API like Exa, gives the assistant the ability to look something up mid-conversation rather than working from memory. I use this when I am fact-checking something time-sensitive, pricing, a recent product change, a version number, and it is the difference between a confident wrong answer and a confident right one.

Developer servers worth keeping

If you write software, GitHub's own MCP server (now maintained by GitHub directly, since Anthropic archived its reference version) is worth the setup for anyone who lives in issues and pull requests. It lets an assistant open, comment on, and review changes without you relaying every detail by hand.

Playwright's browser automation server, maintained by Microsoft, is the other one I would call genuinely useful rather than a novelty. It lets an assistant drive a real browser: click through a flow, check that a page renders, capture what actually loads. I tested it in Claude Code for debugging a broken form submission, and it found the actual failing request faster than I would have by opening dev tools myself.

Both of these are narrow tools for a narrow audience. If you do not write code, skip both.

The categories that sound useful and are not

Generic API wrapper servers for things the model can already do faster by writing code. A calculator server, a unit conversion server, a basic string manipulation server: the assistant does not need a tool call for arithmetic it can already compute or a script it can already write in the time it takes to call an external tool.

Novelty integrations with no repeat use. I connected a weather server, a joke server, and a "quote of the day" server during initial testing. None of them survived a second week, because none of them changed anything I decided to do.

Anything unmaintained. Check the last commit before you install a community server, not after something breaks. The archived reference servers are a useful example: they still work, technically, but nobody is patching them, and an unmaintained server that touches your calendar or your email is a bigger risk than an unmaintained script sitting on your desktop.

Servers that duplicate what your client already does natively. Several clients now have built-in web browsing or file access. Adding a redundant MCP server on top just adds a second permission prompt for the same capability.

Frequently asked questions

How many MCP servers can I connect at once? There is no hard technical limit in most clients, but the practical ceiling is lower than you think. Once you pass six or seven connected servers, tool selection gets noticeably worse, because the model has to pick the right tool out of a longer list on every turn.

Do too many connected servers confuse the assistant? Yes, in a specific way. It is not that the model gets confused about what a tool does. It is that with a large tool list, it sometimes reaches for the wrong one, or skips a tool call it should have made because a similarly named one was higher in the list. Keep the active set small and disconnect anything you are not using that week.

Are official MCP servers better than community ones? Usually, but check which "official" you mean. Anthropic's own reference repository only actively maintains seven servers as of mid-2026. Many categories people assume are Anthropic-maintained, GitHub, Slack, Postgres, were archived and handed to the vendor or a third party. Vendor-maintained (GitHub's own server, Notion's own server) is generally the safer bet over an unlabeled community fork.

Where do I find a trustworthy list of MCP servers? The official registry at registry.modelcontextprotocol.io is the closest thing to an authoritative index, and it launched in preview in September 2025. It lists both official and community servers, so still check maintenance status yourself rather than trusting inclusion in the registry as a quality signal on its own.

Do MCP servers cost money? The protocol itself is free and open. Individual servers vary: many are free open source projects, some wrap a paid API (a search server that bills per query, for example), and a few are offered by a vendor as part of an existing paid plan. Check what is behind the server before connecting it, since the server itself calling an API you already pay for is not an extra cost, but a server calling one you do not have an account for might create a new bill.