A local-first security layer for AI agents that reviews shell commands, file access, packages, plugins, and MCP calls before they run.