WebsitePublisher.ai is a platform that lets AI assistants build and publish complete websites — not generate code, actually publish, on a live domain. Nine AI platforms, 104 integrations, 43 MCP tools. Built in the Netherlands.
This week it got listed in Anthropic's Connectors Directory. Claude users on any plan can connect in one click — claude.ai/directory/websitepublisher-ai — instead of pasting a URL into a config file.
What the thing actually does
43 MCP tools, layered:
Pages, fragments, assets, version history with rollback
Structured data entities and records — so a site can have a product catalogue, a booking table, a member list
Forms, lead capture, visitor authentication via magic link or code
Integrations with an encrypted vault, so credentials never touch the model context
Scheduled tasks
Analytics
An assistant can go from "build me a shop for my ceramics studio" to a live, indexed, form-connected site inside one conversation.
What the review actually tested
Less about features than you'd expect. It was transport correctness at the exact submitted URL, OAuth behaviour, tenant isolation, tool annotations, result sizes, and whether the docs match reality.
Tenant isolation is where a lot of MCP servers will struggle. Every ID-accepting endpoint has to verify ownership against the resolved tenant and return 404 — not 403 — for anything foreign. If you built single-tenant first and bolted multi-tenant on later, you'll find out during review.
One thing nobody tells you
There are two directory surfaces. The public web catalogue at claude.com/connectors, and the in-app catalogue inside Claude. Community-tier listings appear in the second but not the first — so there is no indexable web page for your connector, and no SEO from the listing.
You do get an app deeplink, which is fine for a button but useless as a shareable link: it renders Anthropic's generic card when unfurled, because the route carries no connector metadata.
Worth knowing before you plan the announcement around a landing page that won't exist.
What we'd tell anyone submitting
Test the exact URL you submit, through the public proxy, including every tools/call. Not localhost.
Keep your OAuth metadata endpoints fast.
Make your public docs true. Reviewers read them.
Pick your slug carefully — it's permanent.
Pick your listing name and description for search inside Claude, not for a webpage.
Happy to answer anything about the build or the submission process.
The directory review is a useful signal because it exposes the gap between having MCP functionality and actually being trusted to operate it correctly.
Agreed — though I should temper my own case here.
We're listed community tier, which Anthropic screens but doesn't audit in depth. So the badge is a floor, not a certification. Being trusted to operate it correctly is still mostly on us.
What the review does do is force you to prove the boring parts: transport at the exact URL you submitted, tool annotations that match real behaviour, whether your public docs are still true six months after you wrote them. Those are the things that quietly rot, and in an indie stack nothing else checks them.
So it's less a trust signal than a decay check. Still worth having.