Your first agent

Building your first agent is mostly conversational. You don't click through forms to scaffold one — you describe what you want to Design, and it creates the agent for you. Configure wires it to your install, Review promotes it from draft to live, and then you switch to web and let yourself in so you can chat with it.

Open the dashboard

Run pnpm dev and click the Cast ready at http://localhost:<port> line in the terminal. The dashboard opens.

With no agents on the server, the dashboard docks Design automatically — it opens at a large size in the chat panel at the bottom of the screen. You don't have to find it. Start typing.

Tell Design what you want

Describe the agent the way you'd brief a person. What it's for, who'll talk to it, the tone, anything domain-specific:

Design creates the agent — picks an alias, writes the identity prompt, sets up channels and props, and registers it on the server. A new entry shows up in the sidebar with an amber draft badge. You can keep iterating: ask Design to adjust the tone, add a prop, change a channel. It hot-reloads each change.

Design can also create more than one agent in the same brief. If you describe a team — "an inbox triager and a daily planner that work together" — it'll scaffold both and set them up to talk to each other. For more on that, see Multi-agent composition.

Let Configure wire it up

Design produced the agent's blueprint — what it is. To make it work on your install, Configure handles how it's deployed: model credentials, secrets, transports, extension config. Open All Agents Configure (top of sidebar) or Per-agent Configure (under the agent's row) and tell it what you need:

Notice the password isn't in chat. Configure is a secure surface — no internet egress — so it could take it, but secrets are best kept off the transcript. It opens the credentials page for you instead, then wires the rest as the panels flip green.

Promote it through Review

Drafts bounce inbound messages, so promotion is mandatory before the agent can actually reply. The recommended path is Review, which gives the agent a once-over before flipping it to live:

• On the agent's overview page, an amber "This agent is a draft" banner has a Request review button. Click it.
• Review docks under All Agents and reviews the agent across four lenses — design, configure, economy, security. It surfaces anything risky and asks for your approval before promoting.
• On approval, the draft badge clears. The agent is live.

For your own testing on a throwaway agent, you can skip Review entirely: Settings → Lifecycle → Make live (skip review). The action is recorded in the audit log.

Open the web chat and let yourself in

Now you need to talk to it. The dashboard is the build surface; for actual conversation, every agent has its own web chat URL on the same Cast server. Click the agent in the sidebar — its overview page lists the web chat link.

Open that link in a new tab. The first time, you let yourself in:

1. Send any message in the web chat. Your first message is held, not delivered to the agent yet.
2. Switch to the dashboard. The held message is waiting on the agent's row.
3. Allow it. Choose allow-always, since it's you.
4. You're in. The held message replays and the agent answers.

Talk to it

Type a message in the web chat. The agent replies. That's the success moment.

If something feels off — tone, focus, missing context — switch back to the dashboard and tell Design what you want changed. Changes apply between messages; no rebuild step.

What to learn next

• Access — give other people access on web, Telegram, or Slack (the three transports bundled at launch). Their first message is held until you approve it. Ask Configure to set up whichever transports you want.
• Authoring blueprints — go deeper on what Design just wrote for you: identity, channels, props, the system prompt layers.