- Governed: one policy decides every model call and tool call, and risky actions wait for approval.
- Recorded: every step is in the execution history, so any run can be inspected or forked from any step.
- Isolated: the agent can work in a sandbox that never holds a credential.
- Durable: sessions survive crashes and restarts, and anyone can pick one up from the terminal client.
Quickstart
Run the agent and give it a task.
Terminal client
Work with the agent from a terminal.
Sandboxes
Run the tools in E2B, Daytona, Modal, or Vercel.
Deployment
Run the agent’s image.