Docs
What MCP Atlas is, how to use it, and how it's built — for visitors and for future-me.
What is this
MCP Atlas is a curated directory of real Model Context Protocol (MCP) servers — the tools that let AI coding agents talk to things like GitHub, Supabase, Figma, or your own database. There are a lot of MCP servers now, official and community-built, and it's genuinely hard to tell which ones are real, current, and worth installing versus abandoned or fake.
This directory currently lists 18entries — 16 official and2 community — across6 categories. Every single one was checked live before being added; see how entries are verified.
Using the directory
- Search matches against an entry's name and description, live as you type — no page reload.
- Category filter narrows to one of:
cloud-infra,communication,data,design,dev-tooling,productivity. - Official / community filter — official means maintained by the vendor whose product it integrates with (e.g. Supabase's own MCP server); community means built and maintained by someone else.
- Every entry has a detail page (
/servers/<id>) with the exact install command, links, and a source note explaining how that specific entry was verified.
What each entry means
- Install command
- Copy-pasteable as shown — usually an
npxinvocation, sometimes a manual setup step for remote-hosted servers (Atlassian, Linear) that don't install locally at all. - Maintained
- Whether the entry was actively maintained as of its last-verified date — not a live status check, a point-in-time judgment made during verification.
- Last verified
- The date someone (so far, just the maintainer) actually fetched the repo/npm links and confirmed they resolve. Entries aren't re-checked automatically — an old date is a hint to double check before relying on it.
How entries are verified
The short version: every entry's repo link is fetched directly, and every npm package is checked against the npm registry API (not npmjs.com's package page, which blocks automated requests regardless of whether the package is real). Nothing is added on the strength of a search result alone.
Full methodology, plus the specific catches from the last verification pass (a couple of stale links, one project's own package.json pointing at a dead repo), are on thedata provenance page.
Submitting a server
There's no submission form yet — this is a curated v1, not user-generated. The "Submit a server" link in the header opens a new issue on the official MCP servers repo, which is the right place to propose a genuinely new server to the ecosystem. To suggest an addition to this specific directory, open an issue or PR onthe MCP Atlas repo directly — include the repo/npm link so it can be verified the same way every other entry was.
Architecture
- Astro content collections — entries live in
src/data/mcp-servers.json, validated at build time against a Zod schema (src/content.config.ts). A malformed entry fails the build; it can't ship silently broken. - Zero JS by default — every page is static HTML except the search/filter bar, which is a single small vanilla-JS island. No framework runtime shipped for a directory that's fundamentally just filtering a list.
- No backend — no database, no API, no write path. The whole site is static output, deployed to Vercel.
Running it locally
git clone https://github.com/effinrich/mcp-atlas
cd mcp-atlas
pnpm install
pnpm devNo environment variables needed — there's nothing to configure. See theREADME for the full command list (typecheck, lint, test, e2e).