Guaedclaw (public beta) sits between AI agents and tool access (terminal, files, MCP servers) and blocks risky actions before they run. It is local-first, rule-based, and has 0 LLMs in the security path. 7 defense layers, 1,500+ detection patterns across 11 attack categories. Works with Claude Code, Codex CLI, OpenClaw, Claude Desktop, plus Docker/MCP setups.