Protect local AI coding-agent harnesses before tools run, review approvals and evidence, and scan agent plugins, skills, MCP servers, and marketplace packages with HOL Guard.
Installs into .claude/skills of the current project.
Are you the author of Hol Guard?
Add the live security badge to your README. It updates with every re-scan.
[](https://www.skillsdirectory.com/skills/agentskillexchange-hol-guard)
---
name: "HOL Guard"
slug: "hol-guard"
description: "Protect local AI coding-agent harnesses before tools run, review approvals and evidence, and scan agent plugins, skills, MCP servers, and marketplace packages with HOL Guard."
github_stars: 4
verification: "listed"
source: "https://github.com/hashgraph-online/hol-guard-plugin"
category: "Security & Verification"
framework: "Custom Agents"
tool_ecosystem:
github_repo: "hashgraph-online/hol-guard-plugin"
github_stars: 4
---
# HOL Guard
HOL Guard is a local security runtime for AI coding agents. Use this skill when a user wants to put a supported local harness behind Guard before tools execute, inspect a blocked or approval-gated action, produce receipts or audit evidence, or verify an agent package before release. The `hol-guard` CLI owns harness setup and runtime protection; `plugin-scanner` is a separate scanner distribution for plugins, skills, MCP server packages, and mixed agent workspaces. Do not claim a harness is protected until HOL Guard reports that state, and never bypass a Guard approval. Supported local harness targets include Codex, Claude Code, Copilot CLI, Cursor, Gemini CLI, Hermes, OpenClaw, OpenCode, and Antigravity.
## Installation
No source-backed install or usage instructions could be extracted automatically. Review the upstream project before running this skill in a sensitive workflow.
- Source: https://github.com/hashgraph-online/hol-guard-plugin
## Protect a local harness
Use Guard-owned setup rather than editing agent configuration by hand:
```bash
hol-guard bootstrap
hol-guard install <harness>
hol-guard run <harness> --dry-run
hol-guard run <harness>
hol-guard status
```
For example, use `codex`, `claude-code`, `cursor`, `gemini`, `openclaw`, or `opencode` as the harness name. If Guard queues work, inspect it before approving or denying:
```bash
hol-guard approvals
hol-guard approvals open
hol-guard receipts
hol-guard diff <harness>
```
For diagnostics and evidence:
```bash
hol-guard doctor <harness> --json
hol-guard inventory
hol-guard events
```
## Scan agent packages
Run scanner mode from the package or workspace root so relevant plugin, skill, MCP, and harness configuration surfaces can be discovered together:
```bash
plugin-scanner lint .
plugin-scanner verify . --json
```
Treat scanner failures as real until inspected. Do not mark a package release-ready or a workspace protected without command output proving it.
## Source
- HOL Guard skill and runtime guidance: https://github.com/hashgraph-online/hol-guard-plugin
- HOL Guard: https://hol.org/guard