September 29, 2026
The Claude MCP shortlist
Here’s the uncomfortable truth about the AI you talk to every day: by default it can’t see your codebase, your live website, your documents or your data. It’s a very clever person locked in a room with a keyhole.
MCP (the Model Context Protocol) is how you cut a door in that wall. Each MCP server is a small connector that lets Claude do one real thing: read a page, open a repo, query a database. Pick the right ones and Claude stops guessing and starts checking. Pick badly and you get a cluttered tool list, slower answers and a bigger security surface.
So this isn’t a directory of everything. It’s the nine we’d connect first, and what each one is actually good for.
Disclosure, up front: we build Enrichr, so it appears on this list. We’ve tried to say exactly why, and to be straight about where the others are the better tool.
01 Enrichrour pick
Most “browse the web” tools hand Claude the HTML a server sent. That’s not what you see. Modern sites build themselves in JavaScript, so the interesting parts (layout, motion, 3D, what actually rendered) never appear in a plain fetch. Enrichr loads the page in a real browser and returns what happened: the DOM after scripts ran, the CSS that applied, animations, WebGL, fonts, assets.
Reach for it when the question is about a live site: “how did they build this hero”, “what fonts and tokens do they use”, “rebuild this page”. It’s the only one here focused on that job.
02 GitHub
The official GitHub MCP server lets Claude search code, read issues and pull requests, and see what changed and why. It turns “why is this broken” from a copy-paste session into a question Claude can investigate itself. If you write code, this is the connector that pays for itself in the first hour.
03 Playwright
Microsoft’s Playwright MCP gives Claude a browser it can click through: navigate, fill forms, take snapshots. Where Enrichr tells you what a page is, Playwright is better at proving what it does: does the checkout submit, does the modal trap focus. Use them as a pair.
04 Context7
Models are trained on old documentation, and it shows: deprecated APIs, methods that no longer exist, confident wrong answers. Context7 (from Upstash) feeds Claude current library docs on demand. It’s the cheapest fix we know for “that function was removed two versions ago”.
05 Notion (or your docs tool)
The reason behind a decision almost never lives in the code. It lives in a wiki page, a brief, a meeting note. Connecting Notion or its equivalent means Claude can read the actual spec before it “improves” something you did on purpose.
06 Filesystem
One of the reference servers maintained alongside the protocol itself. It gives Claude scoped read and write access to folders you choose. The keyword is scoped: point it at a project directory, not your entire drive.
07 Supabase
If your app runs on Supabase, its MCP server lets Claude inspect tables, run queries, read logs and check advisories. That’s the difference between “here’s a guess at the schema” and an answer grounded in your real tables. Use a development project while you get comfortable.
08 Sentry
Errors are only useful with context. With Sentry connected, Claude can pull the stack trace, the release and the frequency behind a bug and go straight to a fix, instead of you pasting a wall of logs into chat.
09 Google Drive, Gmail and Calendar
The unglamorous ones that change your week. Claude can find the document you half remember, draft a reply from the real thread, or check your calendar before suggesting a time. Boring, and used daily.