dragon-guard

What problem does it solve?

Prevents agents from executing unsafe or unintended mutating tool calls by enforcing interactive permission checks and mode-based policies so exploratory analysis does not cause accidental changes.

Core Features & Use Cases

• Three-tier modes: Dog (permission-gated), Puppy (read-only planning), Dragon (full implementation) to match user intent and risk tolerance.
• Session-scoped allow/deny lists: Temporarily permit repetitive actions without persisting risky overrides across sessions.
• Safe-tool whitelisting and bash classification: Always allow read-only inspection tools and heuristically classify bash commands to avoid accidental mutations.
• Interactive dialogs and auto-detect: Prompt for permission on ambiguous or mutating actions, optionally auto-switch to Puppy Mode for complex prompts, and summarize risky calls using LLM summaries.

Quick Start

Start in Dog Mode and respond to permission dialogs to allow or block mutating tool calls while you inspect code and plan changes.