- Dormant by default (only wakes on /wake command from owner) - All tools disabled by default (you /enable on demand) - Dangerous actions require /confirm before executing - Single-owner auth (strangers silently dropped, zero info leak) - Full audit log of every action

Built as a working TypeScript prototype with Telegram integration. Next step: connect real AI agent (Claude/GPT) and add WhatsApp support.

Curious what HN thinks about this security model for AI assistants.