With this, your mind can greet a new member with one conversation instead of a catalog of unfinished agents. The page removes navigation choices, not security checks or the agent’s privileges. It is for members of this mind, not public account signup; the public install wizard is a different app with a different authority model.

One fixed thread

apps/onboarding/app.ts exports an app factory like the other member apps. In the publication hook, its install calls itx.agents.create for the fixed /agents/onboarding path. Creating an already existing agent is harmless and creating it does not start a model turn. Every member who opens the page sees the same thread; this is not private storage per person.

// Publish hook, after the Agents runtime is installed:
await onboarding.install(itx);
// Host dispatch:
case "onboarding":
  return onboarding.open?.(request) ||
    this.auth.require(request) || onboarding.fetch(request, this);

index.html and public/app.js provide the chat and public/app.css its styling. The guarded /itx socket is host.auth.itx(request), just as in Agents. The app also answers the shell’s session, one-mind and sign-out endpoints. A tiny shared apps/open-files.ts helper lets an allowlisted icon and web-app manifest load before authentication; it does not make chat assets or actions public.

Instructions are code-adjacent, not pasted forever

The agent’s standing guide is apps/onboarding/onboarding.md. apps/agents/backend.ts selects that file for this agent and reads it afresh for model requests. Updating the guide changes the next answer without duplicating a long prompt into every message. The conversation remains in the agent’s durable context, so a member can leave and return. On the first opening of an empty chat, the page appends a fixed idempotency-keyed context event asking the agent to begin. That is distinct from installation: publishing creates the agent without starting it, while opening the empty chat starts the greeting once. The Dash can open this same agent, so preserve a shared first-open contract instead of manufacturing a fresh conversation per entry point.

The guide also directs account connection through iterate Cloud and asks the agent to verify the resulting vault entry and one real provider response. Consent pages and passkey approvals stay on platform surfaces, not in chat-built credential forms. A conversational yes is not a substitute for a required signed approval. In the publication hook, independent install steps report their own failure without preventing the later apps from initializing; still check that this fixed agent actually exists before claiming onboarding is ready.

Adapting the pattern

Install the normal Agents runtime, choose a fixed agent path, point its standing files at a guide of your own, and build the focused member page using the same authenticated socket contract. This needs the mind’s normal model funding, not a new vendor account just for onboarding. Test repeat publication, an empty thread, a returning member, a signed-out visitor and a signed-in nonmember. Do not use an ordinary full-mind onboarding agent as an isolated public helper: the install wizard explains why its end-user chat has a jail and bounded tools instead. The blog keeps this recipe current when the app’s code changes.