Which CLI runs the work

Flowviant doesn’t run its own AI. Every agent is one of the coding CLIs — Claude Code, Codex, or Antigravity — running on your machine, signed in with your own account for that tool. Which CLI an agent uses is decided once, when you accept it in the plan split, and it stays that CLI for as long as that agent exists: an agent’s value is the conversation it’s holding, and that doesn’t carry over if you switch tools mid-work.

When you plan a stack of cards, a proposed agent shows a CLI picker if your machine reports more than one CLI it can drive — pick one before you accept. A design or research card always runs on Claude Code.

If you don’t pick one, the project falls back to a default. Project Settings → Agents lets you set a default CLI for code cards, and a fallback order for when there’s no default. If none of that resolves to something available, Flowviant falls back to whatever CLI the machine leads with.

The machine and its CLI logins are shared: everyone working on a project uses the same computer’s logins. If a CLI hits its own usage limit, every agent running on that CLI is parked, and the daemon says which CLI hit the limit and relays that CLI’s own reason. A parked Codex agent offers Run on Claude instead on its page — pressing it switches that one agent to Claude Code. It keeps the agent’s branch and its cards, not Codex’s own conversation history, so it picks up where the branch left off rather than where the chat left off.