Several AI agents work on your idea at the same time. Work moves across a task board, and every finished task is scanned for leaked secrets and unsafe code before it reaches you. The merge stays your call. You also get a chat with each agent.
The director splits the mission into cards and dispatches them to coder, researcher, and auditor seats - each in its own pane, on its own model.
An auditor seat reviews each builder’s finished work, and every finished card is scanned for secrets, unsafe code and CSP problems - findings are bannered, never auto-passed, and the merge is yours. Footguns like bypass and broadcast stay loud in the chrome.
Agents hand off worktrees and card references - never re-stated context. Token spend stays flat as the team grows, and every handoff lands in the Vault.
An agent that finds its card too big spawns a swarm underneath it, scoped to its own directory. The tree shows which run belongs to which parent, so a nested team is never an orphan.
Hourly and lifetime token caps per role, warned before they bind. Pause holds the whole roster and shows you which seats have acknowledged it - a paused run says so, it does not quietly finish the card in flight.
Save a roster that worked and spawn it in any workspace. Afterwards, the router's dispatch timeline replays so you can answer why the coordinator sent a card where it did - without re-running the swarm to find out.
Mix providers inside one swarm: Opus on the hard pane, a cheap model on the researcher, a custom endpoint on the auditor. On keys you bring, Orcyx adds no markup, and per-pane spend stays in the chrome where you can see it.
Live recreations of two screens: the board that shows who is working, and agents reporting back side by side. Hover over them.