Nova

Nova is ClusterCode’s automation agent — the one that actually runs a Run or a Schedule. It works on your behalf in the background: it reads your job prompt, picks a host, launches a DevBox, and delegates the real coding to an in-DevBox engine, then reports the result. It’s the coordinator in the two-layer model — distinct from the engine that writes the code.

Nova runs on Anthropic models

Like Nucleus, Nova’s coordination is powered internally by Anthropic (Claude) models, drawing on your plan’s AI allowance. To add usage beyond your plan — or run on your own account — bring your own key in Settings → Secrets: an Anthropic API key or a Claude OAuth token. (The in-DevBox engines may use their own credentials — see Engines.)

Nucleus vs Nova

ClusterCode has two agents, both Anthropic-powered, with different jobs:

• Nucleus — the interactive assistant in the dashboard. You chat with it to manage your cluster (launch DevBoxes, run commands, check status).
• Nova — the background automation agent. You don’t chat with it directly; you give it a job (a Run or a Schedule) and it executes that job end-to-end, on its own.

Two modes

Nova is the same agent in both cases — only the contract changes:

• Interactive — a Run — a human started it and is watching live. Nova can pause and ask you a question, and you can steer it through the chat. This mode owns the “ask the human” tool.
• Autonomous — a Schedule (or an autonomous Run) — headless, on a cron cadence, no human in the loop. There is no chat, no clarification, no follow-up turn: Nova executes the job prompt once and reports a result. Anything ambiguous it decides for itself.

Note

Because the autonomous mode can’t ask you anything, a Schedule (and an autonomous Run) must be given its host and image up front. See Runs → Interactive or autonomous.

What it does each run

Whichever mode, a Nova execution follows the same arc:

1. Read your instruction (and any ticket context).
2. Launch a DevBox on the chosen worker or cloud worker.
3. Delegate the coding to an engine (Claude Code, Codex, or Copilot) inside the DevBox, and run quick checks itself.
4. Track every DevBox it touches, so the orchestrator knows what was created even if the run dies mid-job.
5. Report the result — a completed/failed summary, a PR link, filed issues — on the timeline.

Next

• Run an agent on demand — recipes for interactive Runs.
• Automate recurring work — recipes for Schedules.
• Engines — the CLIs Nova delegates the coding to.