Get started

Add a model

Add a provider in Settings → Model provider and paste your API key. Endpoints for the common vendors are pre-filled.

The key is encrypted on this machine and sent only to that provider.

Open a project

Add a folder under Projects in the sidebar and the agent works there: its file tools stay inside it, and commands need your approval by default.

You can also chat without a project — the agent uses a default workspace, which you can change in Settings → General.

Your first conversation

Describe the task as you would to a colleague — “add a remember-me option to the login page”. The agent reads the code first, then proposes changes.

Type @ to reference a file, / to use a skill.

Approvals and rollback

Before a write or a command, an approval card appears. Allow lets this one through; Always allow stops asking for that kind of action in this project. Commands can never be allowed for good.

Switch the approval mode next to the composer: Ask each time, Auto (safe) or Auto (all). Auto (all) skips every confirmation — use it only when you fully trust the task.

Every turn can be rolled back in one click.

Plan mode

Press ⇧Tab. The agent researches read-only and proposes a plan; nothing changes until you approve. Good for larger changes.