Get started

Add a model

Open Settings → Model provider, pick Anthropic, OpenAI, DeepSeek or a custom OpenAI-compatible endpoint, and paste your API key.

The key is encrypted on this machine and sent only to the provider, by the app, when it calls the model.

Open a project

Add a folder under Projects in the left sidebar. Conversations started in that project work inside that folder: the file tools can’t reach anything outside it, and running a command always needs your approval.

You can also chat without a project; the agent then works in a separate default folder, which you can change in Settings → General.

Your first conversation

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

Type @ in the composer to point at a file, or / to use a skill.

Approvals and rollback

Every file write or command shows an approval card. Allow lets this one through; Always allow stops asking for that kind of action in this project (running arbitrary commands can never be allowed permanently).

The approval mode next to the composer has three levels: Ask each time, Auto (safe) and Auto (all). Auto (all) turns confirmations off — use it only when you fully trust the task.

Every finished turn has a Revert this run button in the conversation.

Plan mode

Press ⇧Tab. For that turn the agent only reads, then hands you a plan; it starts changing things only after you approve. Good for larger changes you want to agree on first.