Skip to main content
Creating an agent asks for a name. That is the entire form. Everything a wizard used to ask for — skills, channels, browser mode, the agent’s personality — is configured inside OpenClaw afterwards, by the person who will use the agent. See Inside the agent.

What happens when you press the button

1

A seat is claimed

If your plan has a free seat, it is taken. If it does not, the screen shows a plan chooser instead of the form — with your typed name kept — and checkout returns you here.
2

A provider key is issued

The agent gets its own model key with a spending limit. Until the first deploy the agent has no key at all, which the balance panel states plainly rather than showing a misleading $0.00.
3

A container and a volume are provisioned

One virtual machine and one block volume, for this agent alone. The volume outlives restarts, redeploys and version upgrades.
4

OpenClaw starts inside it

The platform writes a starting configuration once, on first boot, and never again. From that point the config file belongs to the agent — nothing on the platform overwrites it.
Under five minutes, end to end. The page shows an estimate the platform refines from its own measured runs, so it gets more accurate as your install accumulates history.

The budget line

Before you create, the screen states what the agent starts with — and zero is a real answer that gets said out loud: “No budget yet — it will start and wait.” An agent with no money still deploys and still runs. It cannot make a model call until there is something to spend, and it will sit quietly rather than failing.

Created, but not started

Two outcomes look like failure and are not the same:
Capacity is finite. The agent exists, its seat is paid for and stays yours, and starting it again costs nothing. The page offers to retry, and a queue position is shown while it waits.
A failure notice appears above everything else on the agent’s page, classified where the container said enough to classify it — configuration, image, or runtime. It carries a contact-support action with the detail pre-filled.

Naming

Up to 100 characters. The name is what appears in the agents table, in the Telegram Hub agent picker, on alerts, and in the seat ledger. It can be anything; it is not an identifier.
Last modified on August 31, 2026