// How it works
From a CLI to a fleet of agents, in five steps.
Configure the CLI, point it at your stack, prove quality with evals, and deploy. From there the Command Center runs the whole fleet — every agent observable, recoverable, and updatable.
Install the CLI.
Point it at your stack.
One CLI is the control surface for everything that follows. Install it, log in, and register the endpoints your agents authenticate and talk through — the same endpoints they'll use in production.
Wire up data, tools,
and the agent itself.
An agent is its instructions, its skills, and the data it can reach. Define all three in version-controlled files — global data shared across the fleet, per-user data scoped to each subject, plus the skills and AGENTS.md that shape behavior.
Prove it before
it ships.
An eval is a prompt plus what you expect back. Check the output objectively with a grep, subjectively with an LLM judge, and assert the exact tool calls the agent should make. Run it against faked MCP fixtures or your real MCP — your choice, per eval.
One line.
The whole fleet.
When the evals pass, ship. A single command provisions every agent warm, opens a two-way channel to your backend, and hands you a live dashboard — no queue, no cold start, no per-agent setup.
The operator layer for thousands of persistent agents.
Once agents are live, the Command Center is where you run them — the control plane that makes a persistent fleet observable, recoverable, and updatable, without treating each agent like a bespoke snowflake.
Five steps from your stack
to a fleet you can run.
Configure, set up, prove with evals, deploy — and operate every agent from one console. See it on your data, with your tools.