Start here

You are talking to the catalog of a lab about cross-boundary agent trust — the problem of letting an agent work with systems and other agents that belong to somebody else, without either side taking on risk they cannot account for.

Most enterprises solve this today by not solving it. The Lab is a neutral space, outside anybody’s perimeter, for finding out what is actually possible now.

What is in here

Setup steps — Lab infrastructure you connect to or configure once: an account, a signed-in client. You do not run these; you get them in place, in order, and then forget about them.

Recipes — guided, runnable paths. Each clones from a GitHub repo and runs on your own machine against the Lab’s hosted infrastructure. Each says what it teaches, what it needs, and roughly how long a first run takes. This is where the learning happens.

Targets — sample MCP and A2A endpoints the Lab hosts for your agent to call during a recipe. Most admit only the Agent Gateway, so you will reach them through a governed path rather than directly. That is the demonstration, not an obstacle.

If you are new

1. Work through the setup steps in order — get_setup_step for the detail of any one of them.
2. Then list_recipes with difficulty: introductory and status: available.
3. get_recipe gives you the repo, the entry point, and what to expect.

Not everything listed is built. status tells you the truth: available means you can clone it and it works today. planned means the shape is agreed and the work is not done. Nothing is marked available on the strength of an intention.

What the Lab expects of you

Bring your own compute for agents you run, and your own model API keys. Do not load real or sensitive data — this is an experimental environment with no warranty and no SLA. Build something worth showing, and take it back to your organisation as evidence.

If you are stuck

Every recipe that calls a Gateway-fronted target explains how to read Payload Capture, which shows all four stages of a request. Start there — you will not be able to call those targets directly to isolate a problem, by design. The community Slack is the next stop.