hashgraph-online/hol-guard

Open-source antivirus for AI agents: block risky tools, secret access, prompt injection, malicious packages, MCP servers, plugins, and skills at runtime.

What it solves

HOL Guard is a local-first security layer designed to protect machines from risky actions taken by AI agents. It acts as an "antivirus" for AI, preventing agents from exposing secrets, executing unsafe shell commands, falling victim to prompt injection, or installing malicious packages and MCP servers.

How it works

Guard integrates with various AI agents (such as Claude Code, Cursor, and GitHub Copilot CLI) using native hooks, managed proxies, or launch overlays. It evaluates agent actions and local artifacts against a security policy before they are executed. Depending on the policy level (Gentle, Balanced, Strict, or Paranoid), Guard can allow safe actions, block known threats, or pause ambiguous actions for human approval via a local dashboard or terminal.

Who it’s for

It is built for developers and security-conscious teams who use AI agents and developer tools to automate tasks on their local machines and want to ensure these agents do not compromise their system or leak sensitive data.

Highlights

  • Multi-Agent Support: Integrates with a wide range of agents including Codex, Claude Code, Cursor, Gemini CLI, and Grok.
  • Runtime Protection: Screens tool calls, shell commands, and file access in real-time.
  • Supply Chain Security: Reviews plugins, skills, and MCP servers before they are trusted.
  • Llocal-First: Works locally without requiring a cloud account, though an optional Guard Cloud can be used for team policies and synchronized evidence.
  • Human-in-the-Loop: Provides a local approval center to review and authorize paused actions.
  • Command Inspection: Includes tools to test and explain command safety without actually executing them.