
When you start a conversation, you can pick a permission level for the run, controlling exactly what the Agent is allowed to do and balancing safety with efficiency.
There are four levels: Read-only — the Agent makes no changes to external systems through Connectors or Castrel Proxy nodes, but can still record findings and write to the knowledge base inside Castrel; Request Approval — every outbound write asks for your confirmation first; Approve for Me — the AI clears routine actions automatically and only stops for your sign-off on risky ones; Full Access — no manual approval at all, for fully trusted automation.
Castrel now supports theme switching — move freely between dark and light themes and adjust the font size for the interface and chat. Chat, tasks, documents, and the terminal all follow your settings, so readability stays consistent under any theme.
Appearance settings now follow your account. After login they load automatically, save the moment you change them, and restore on refresh, re-login, or a new device — no longer tied to a single browser. When you're signed out or the server is unavailable, the interface falls back to local settings.