Agent modes

Choose a model, set reasoning effort, and toggle Plan and Fast modes.

Every chat session carries its own model, reasoning effort, and execution modes. You set them in the chat input, and Kalauz remembers them per session.

Models

Kalauz routes to whichever agent CLI owns the model you pick — Claude or Codex. Models whose CLI isn't installed are greyed out.

The 1M variants carry a one-million-token context window. New workspaces default to Opus 4.7 at high effort.

Reasoning effort

Effort controls how much the model deliberates before acting: low, medium, high, or max. Higher effort trades speed for thoroughness — reach for it on gnarly refactors, drop it for quick edits. (Codex tops out at high, so max maps to high there.)

Plan mode

Turn on Plan and the agent runs read-only: it investigates and proposes a plan before touching anything, and asks permission for tool use rather than acting freely. Good for unfamiliar code or high-stakes changes where you want to approve the approach first.

With Plan off (the default), the agent works autonomously in its isolated worktree — safe precisely because the worktree is sealed.

Fast mode

Fast favors lower latency for snappier back-and-forth on smaller tasks. Like the other settings, it's remembered per session.

Switching mid-task

Change the model, effort, or modes at any point — the next turn uses the new settings. Kalauz tracks Claude and Codex sessions separately, so switching provider resumes the right conversation rather than starting over.

What's next?

• Configure model providers — connect and authenticate Claude or Codex.
• Agent behavior — the exact rules a run executes under.