Connecting an MCP Server We've Never Heard Of

Until now, connecting a Model Context Protocol (MCP) server to your agent meant picking one from our catalog — a curated list of about 64 known servers. If the tool your team actually needed wasn't on that list, there was no way in. That's changed: you can now register any public MCP server by URL, and your agent can start using its tools the same way it uses a catalog server's.

Why this matters

MCP has grown fast enough that there are now thousands of public servers, and the reason most teams want to add one isn't curiosity — it's that their data lives somewhere no public catalog reaches. An internal CRM, a proprietary system, a vendor's own MCP endpoint. Waiting for us to review and list a server before you could use it was a real bottleneck, and it's gone.

What we checked before opening this up

A custom URL is a wider door than a fixed catalog, so we validate it more than once, not just on the way in. The server's URL is checked before it's saved, checked again when it's read back, and checked a third time right before every connection attempt — specifically because a redirect can point somewhere unsafe after the URL was already approved once, and checking only at save time wouldn't catch that.

One deliberate limit

Custom servers work over HTTP only — we don't support stdio transport, and we're not planning to. stdio expects a local process and a filesystem to launch it in; a Workers-based platform like ours doesn't have either, so a server that only speaks stdio isn't something we can connect to no matter how it's added. If a server you want to use only ships a stdio client, it needs an HTTP-reachable option to work here.

Custom MCP servers are available now in AI Agent's connection settings.