Most catalogs assume a person with a browser. This one does not — the Lab publishes its own catalog over MCP, so the first thing your agent learns is how to find out what else it could be doing.
What you do
Give an MCP client one URL. Claude Code, Claude Desktop, your own agent — anything that speaks the protocol.
It comes back and tells you to sign in. A browser opens, you authenticate as yourself, and the client has what it needs. Ask your agent what recipes are available and it answers.
What you did not do
You did not create a second account, generate an API key, or paste a secret into a configuration file.
Your client asked the server where to authenticate, ran the flow itself, and now holds a short-lived token that represents you. Revoke it centrally and every agent holding it loses access at once.
Most agent tooling today works the other way round: a long-lived key with no subject, no expiry, and no way to revoke one holder without revoking all of them. That difference is the reason the rest of the Lab is worth your time, and you have just met it before reading a word about how it works.
Then do it again through the Gateway
Change the URL to the Gateway’s and run the same thing.
From your agent’s side nothing has changed — same tools, same answers. What changed is that a policy decision now sits in front of the catalog, and a credential your agent never sees is attached on the way through. If the decision goes against you, you get a refusal instead of an answer, and reading that refusal carefully is the more instructive outcome.
This is the safest place in the Lab to meet a trust decision for the first time. The thing on the other end is a read-only list of documents, so a misconfiguration costs you nothing except the lesson.
Where next
Ask your agent for introductory recipes, or for anything exercising a part of the Trust Fabric you have not met yet. From here on you should not need to read a web page to find your way around — that is rather the point.