Chat with an agent on a machine

Opening a computer, or the Agent section of a sandbox, opens one chat bound to that machine. The workspace panel sits beside it.

Pick the brain in the chat

Before the first message you choose what powers the chat:

  • Optimal runs a machine-bound chat on wallet credits.
  • A harness runs on your own account: Claude Code, Codex, OSA, pi, OpenCode, Prime Agent, Kimi Code, or Mistral Vibe.
  • A saved agent runs with its own instructions, tools, and permissions.

The chat is bound to the machine you opened, so there is no “where should you build?” step.

The brain is remembered per person, per workspace, and per machine. It never changes the brain you picked on the Optimal page.

While it works

  • Switching the brain stops the run, keeps the old thread above a divider, and starts a new thread.
  • A Connect card opens when your account is not connected.
  • Approvals appear inline, and you can approve or deny.
  • Stop cancels the run.
  • Take over hands control to you, and Hand back returns it.

The panel follows the machine

The workspace panel opens on Desktop for a computer and on Terminal for a sandbox, then Preview once a URL exists. It restores the layout you had on the Optimal page when you leave.

A machine that belongs to another workspace never gets a chat, a chat id, or a run.

Was this page helpful?