Show HN: SafeClaw – Sleep-by-default AI assistant with runtime tool permissions (github.com)

SafeClaw, a new personal AI assistant, has been introduced with a focus on user control and privacy. Unlike traditional AI assistants that operate with continuous access to user data and tools, SafeClaw starts in a "sleep-by-default" mode, activating only when the owner explicitly sends a command via Telegram (and soon WhatsApp). This unique approach addresses significant trust issues prevalent in existing AI platforms, where potential security vulnerabilities may arise from always-on functionalities and excessive permissions. The technical structure of SafeClaw emphasizes a single-owner model and runtime permissions, where all six categories of tools (browser, filesystem, shell, etc.) are disabled until manually enabled, mitigating risks linked with unsolicited access. Each action taken by the user requires confirmation, creating an added layer of security. The assistant features comprehensive audit logging and a dormant state that automatically reverts if inactive for 30 minutes. While currently functioning as a prototype with simulated tools, SafeClaw plans to integrate real AI capabilities and expand tool management features, highlighting its potential to enhance user safety in AI interactions.