The safety layer between AI coding agents and your shell. We make guard, a free, source-available check on the shell commands Claude Code and GitHub Copilot CLI want to run.
Boxdawn makes safety tooling for AI coding agents. Agents like Claude Code and GitHub Copilot CLI run shell commands on your machine, and one wrong line can delete files or throw away work. We make guard: before a command runs, it denies it, asks you with the reason, or leaves no decision to your own permission settings. It never answers "allow". guard is free and source-available at github.com/boxdawn/guard. We publish what it decides on commands quoted in public incident reports, including the ones it says nothing about. Founded in 2026 in South Korea.
Our stack
Boxdawn guard: Python 3.12; tree-sitter with tree-sitter-bash to parse Bash; on Windows, PowerShell's own parser (System.Management.Automation) through a small C# helper; cryptography to check signed rule files. Built with PyInstaller into one standalone executable per platform: Windows x64, macOS (Apple silicon and Intel), Linux (x64 and ARM64). Runs as a Claude Code plugin hook (PreToolUse) and a GitHub Copilot CLI plugin hook (preToolUse). Tested with pytest on GitHub Actions; each release runs real Claude Code sessions on all five platforms. License: FSL-1.1-ALv2.