Get started

Try it in the console

Everything the quickstart did from a terminal, the console lets you watch and repeat: the deploy as it runs, a conversation with the agent, what each run did, and what it cost.

Find your project

Sign in at /app/home. The first time, the console asks you to turn on a second factor before it shows anything else. It then takes you into a project.

The rail on the left has two halves. Above the divider is one project — the picker at the top chooses which, and its All projects… entry lists every project in your organization. Below the divider is the organization as a whole. On this page, a path starting with Build, Productize, Run or Settings is in the project half; one starting with Billing or Config is in the organization's.

Watch the deploy

Build → Deployments lists the project's deploys, newest first. Open one and you get its state, the phase it is in while it runs — the page refreshes itself until the deploy ends — and two tabs:

• Phases: the pipeline from validate to verify, marking where a failed deploy stopped, and what each target runs now.
• Scan: what the image scan found, counted on the tab — or that the scan did not finish, which is not the same as a clean image.

A deploy refused before anything was built lists every problem at the top of the page. One that failed after provisioning says that resources may have been left behind.

Talk to it in the Playground

Open Build → Playground, or the Launch in Playground icon on the agent's row under Build → Agents. The Playground takes the whole window; Back to your project's name returns you to the console. Choose the agent, press New conversation, and type.

• The composer waits until the indicator says Live. A reply travels over an event stream that has to be subscribed before the message is sent, or the reply is lost.
• Your first message creates the conversation and becomes its title. Conversations are listed on the left, read from the agent's own database, so reopening one shows exactly what was said — tool calls included.
• It runs as you, billed to your organization like a run from the CLI. No service user and no end-user wallet is involved.
• Agents only, for now. A workflow is launched with typed inputs rather than a chat, and the Playground says Workflow support is coming soon instead of offering a composer.

Attach a file

The paperclip (Attach a file) uploads a file to this conversation before your message is sent, and the Playground adds (attached: report.pdf) to the message for you. That name is the attachment: the platform hands the model the bytes of a file a message names. Files and multi-modal covers what a model can read and where a file lives.

Answer a pause

When the agent stops to ask, a card titled The agent is waiting appears in the transcript. What it offers depends on what the run is waiting for:

The shield icon beside the agent's name opens Action settings: what each of its declared tools does for you — acts, asks or is refused — and why. A grant you made with Always approve is listed there with Ask me again to take it back.

Run options

The sliders icon beside the composer opens Run options for the next message: session state to merge into the conversation, and a model call limit, which can only lower the agent's own.

Read what it did

Run → Observability answers “is anything wrong right now”. Pick a window and, optionally, agents or workflows or one agent, and it shows the window's headline figures, the activity mix, run duration, tool latency (p95 by tool), and the failures among the most recent 200 runs, each with its error.

For one run in full, open Build → Agents, the agent, and its Traceability tab, which lists its runs. A run opens on a page of its own: duration, tokens, model and tool calls, the execution tree, every step's detail, and Download JSON for the raw events.

What it cost

Run → Analytics reads a time range — 24 hours, 7 days, 28 days (the default), 90 days, 6 months or Custom — through three tabs:

• Overview: what the project earned from your own users' wallets, and its platform cost — compute, storage, builds, approvals, wallet calls and the rest, priced by a rate card you can read and cannot set. Model tokens are not on it: you bring your own key, so your provider bills those.
• Tokens: token counts, per agent and version.
• Runs: how many ran and how fast, agents and workflows apart — a workflow's duration includes waiting for a person.

Across the whole organization, Billing → Usage (/app/usage) has the platform charges, your balance, account credit and payments, and Billing → Invoices what has been billed. The Billing group is shown to organization admins only. Billing and the ledger explains the numbers.

Give your backend a credential

Settings → API access issues the credentials your own backend uses — a client id and secret it exchanges for a token, not a key it sends on every call. New service user asks for a name, an optional description, and whether it Can write: on by default, and needed to credit a wallet or mint a ticket. The secret appears once, beside the token exchange already filled in, and nothing can show it again; Rotate issues a new one and stops the old, and Remove deletes the identity. Only organization admins see this page's list.

Call it from your code picks up from here.

Where everything else lives