Plan and Code modes

The toggle beside the prompt box decides how much the agent does before it checks in with you.

Plan

The agent reads files, searches the codebase and asks questions — but does not edit anything. It ends by proposing an approach for you to approve, redirect, or throw away.

Use Plan when:

• the change touches code you do not know well
• there is more than one reasonable approach and you want to choose
• the task is large enough that a wrong start wastes real time
• you want to understand the codebase before changing it

Code

The agent works through the task, editing files and running commands as it goes. It still asks before anything destructive, but it does not stop to confirm each step.

Use Code when:

• you can describe precisely what you want
• the change is contained, or you have good tests
• you are iterating on something the agent just did

Switching mid-task

Toggle at any time. The agent keeps the conversation and everything it has learned — switching from Plan to Code after approving an approach is the common path, and the plan it just wrote becomes its instructions.

Different models per mode

Under Settings → Model, enable Use different models for Plan and Code modes.

This lets you run a strong reasoning model for planning and a cheaper one for execution. Planning is where model quality pays off most; carrying out an approved plan is often work a smaller model does perfectly well.

Each mode then remembers its own model, and switching modes switches models with it.

What this is not

Plan mode is not a safety boundary. It does not edit files, but it still reads your code and sends it to the model — the same as Code mode. If you need to restrict what the agent can touch, use a .toolsxpoignore file rather than relying on the mode.