panguard-ai/panguard-ai

---

Your AI agents now act on your behalf. They install skills from registries you’ve never read, and they run shell commands, edit files, and fetch URLs on their own. Someone should check what they’re about to do.

PanGuard is that check. One command scans every skill your agents installed against 747 open ATR rules, drops a guard between your agents and the tools they call, and shows you the whole thing in a local dashboard. Nothing about your code leaves your machine.

npm install -g panguard && pga up

---

Why I built this

I’m Adam Lin (林冠辛), a solo founder in Taiwan. I’m not a career engineer — I came from real estate sales and marketing, ran a hip-hop festival for five years, and taught myself to code. When AI agents started installing skills and running tools on my own machine, I realized there was no antivirus for any of it: a skill is just text an agent will follow, and a malicious one can tell your agent to exfiltrate your keys or curl | bash a payload — and nothing was watching.

So I built the thing that watches — and the open detection standard underneath it, ATR, the way Sigma standardized SIEM rules and YARA standardized malware signatures. ATR is independent and MIT-licensed; PanGuard is one runtime on top of it.

Solo, from zero — and the rules behind it are already merged into Microsoft’s and Cisco’s security tooling (details below). I’m giving the whole runtime away free because this should exist for everyone, not sit behind a sales call.

---

What you get

• Pre-install scan — pga scan audits a skill or MCP server against 747 ATR rules in seconds, and exits non-zero on a real threat so any CI can gate a malicious skill out.
• Runtime guard — a local daemon inspects every agent tool call — both MCP servers and the agent’s own built-in tools (Bash / Edit / Write / WebFetch) — across 7 platforms (Claude Code, Cursor, Codex, Continue, Gemini CLI, Cline, Windsurf), and blocks known attacks before they run.