Security guard hook for Claude Code. Evaluates every tool call (Bash, Edit, Write, Read, MCP) against 864 pattern-matching rules to block dangerous commands, data exfiltration, and supply chain attacks. Zero external dependencies, written in Zig. Supports three decision levels: allow, deny (hard block), and ask (user confirmation prompt). Protects against reverse shells, pipe-to-interpreter execution, credential leakage, cloud metadata access, environment variable injection, and more.