Engines

Every Schedule and Run has two layers. The engine is the inner one — the CLI that does the actual coding inside the DevBox.

Two layers: coordinator and engine

• Coordinator — the agent that runs the job: it reads your instruction, picks a host, launches the DevBox, and delegates the coding. In an interactive Run the coordinator owns the “ask the human” tool, so it can pause and ask you a question regardless of which engine you pick. (A Schedule — or an autonomous Run — has no human, so it never pauses.)
• Engine (subagent) — the CLI that writes the code inside the DevBox. You choose it per Schedule or per Run.

The engines

Engine	Billing	Pauses to ask you?	Notes
Claude Code — Plan	Your weekly plan limit	Yes, natively — it can stop mid-task and surface a structured question on its own	Recommended for interactive / exploratory work
Claude Code — API credits	Your monthly API credit pool	Yes — the coordinator relays its question	Runs a turn straight through
Codex	Your OpenAI / Codex auth	Yes — the coordinator relays its question
GitHub Copilot	Your Copilot auth	Yes — the coordinator relays its question

Note

Can every engine pause and ask? In a Run, yes. The difference is how: Claude Code — Plan can pause itself mid-task with a structured question, while with the others the coordinator asks on their behalf (for an ambiguous instruction, a host/image choice, or to relay a question the engine raised). In a Schedule or an autonomous Run, none of them ask — there’s no human in the loop.

The two Claude engines

They run the same Claude Code CLI; the difference is how the AI is billed and how it pauses:

• Claude Code — Plan bills against your weekly plan limit and can pause itself mid-task to ask you something. Best when you want a back-and-forth.
• Claude Code — API credits bills against your monthly API credit pool and runs each turn to completion. Best for unattended, well-specified work.

Choosing an engine

• Auto-detect (the default) picks the best available CLI on the image.
• Name an engine explicitly when you care which one runs — e.g. an interactive brainstorm is a good fit for Claude Code — Plan, while a tightly-scoped unattended job suits Claude Code — API credits or Codex.

See Runs for how a paused question looks, and Schedules for the autonomous contract.