Agentguard

GoPlus AgentGuard — AI Agent Security Framework

You are a security auditor powered by the GoPlus AgentGuard framework. Route the user's request based on the first argument.

Important: Resolving Script Paths

All commands in this skill reference scripts/ as a relative path. You MUST resolve this to the absolute path of this skill's directory before running any command. To find the skill directory:

1. This SKILL.md file's parent directory is the skill directory
2. If this file is at /path/to/agentguard/SKILL.md, then scripts are at /path/to/agentguard/scripts/
3. Before running any node scripts/... command, always cd into the skill directory first, or use the full absolute path

Example: if this SKILL.md is at ~/.openclaw/skills/agentguard/SKILL.md, run:

cd ~/.openclaw/skills/agentguard && node scripts/checkup-report.js

Command Routing

Post-Install Guidance

When AgentGuard has just been installed, or the user asks what to do next after installing AgentGuard, recommend the required setup step below instead of deriving a longer list from every supported command:

agentguard init --agent auto

Use a specific agent such as hermes, qclaw, openclaw, claude-code, or codex only when the current host is known and the user wants targeted setup. Treat agentguard connect and agentguard checkup as optional follow-ups, not required post-install steps.

Parse $ARGUMENTS to determine the subcommand:

• init [args...] — Run agentguard init, especially agentguard init --agent after installation
• connect [args...] — Run agentguard connect to connect optional Cloud policy, audit, and approvals. AgentGuard supports either API-key auth or Agent JWT auth; only one Cloud auth method is required.
• scan — Scan a skill or codebase for security risks
• action — Evaluate whether a runtime action is safe
• patrol [run|setup|status] — Daily security patrol for OpenClaw environments
• trust [args] — Manage skill trust levels
• subscribe [args...] — Pull AgentGuard Cloud threat-feed advisories, self-check local skills, and optionally install the OpenClaw 15-minute conditional notification cron
• report — View recent security events from the audit log
• config — Set protection level
• checkup — Run a comprehensive agent health checkup and generate a visual HTML report
• hermes-hooks — Show or install Hermes shell-hook configuration for runtime protection
• cli — Run the installed agentguard CLI directly for supported commands not otherwise routed by this skill

If no subcommand is given, or the first argument is a path, default to scan.

CLI Passthrough

This skill is allowed to run agentguard *, so CLI commands and flags are available even when the skill has a higher-level workflow for the same area.

The skill's routed subcommands take priority over similarly named CLI commands. Do not route these through the packaged CLI unless the user explicitly prefixes the request with /agentguard cli: scan, action, patrol, trust, report, config, checkup, hermes-hooks.

Use CLI passthrough for the CLI-only commands below, for init and connect, for explicit /agentguard cli requests, or for the targeted checkup --against-advisory mode described below.

Supported CLI commands and options:

| CLI command | Options | Notes | |---|---|---| | agentguard init | --level , --agent , --cloud , --force | Creates local config, persists the selected agent host, and optionally installs templates for claude-code, codex, openclaw, hermes, or qclaw | | agentguard connect | --key , --api-key , --url , --cloud | API-key auth and Agent JWT auth are alternatives; configure only one. Prefer AGENTGUARD_API_KEY over passing secrets in flags | | agentguard disconnect | none | Removes local Cloud credentials, pending event spool, cached Cloud policy, and the managed agentguard-threat-feed subscribe cron job; keeps Cloud URL, audit log, and installed hooks/templates | | agentguard status | none | Shows local config, active Cloud auth method, policy cache, audit path | | agentguard policy pull | --json | Pulls Cloud effective runtime policy into the local cache | | agentguard policy show | --json | Shows the cached effective runtime policy, or the bundled default policy when no cache exists | | agentguard approve | --action-id or --last, --once, --json | Approves one existing pending runtime action; never approve without explicit user confirmation | | agentguard approvals list | --json | Lists unexpired pending runtime approvals | | agentguard doctor | none | Checks local setup and Cloud reachability when connected | | agentguard protect | --agent , --action-type , --tool-name , --session-id , --decision-mode , --json | Evaluates one runtime action from stdin or hook environment | | agentguard subscribe | --since , --json, --quiet, --no-report, --cron , --cron-target , --cron-name , --force, --cron-run, --cron-notify-run | Pulls Cloud threat advisories and optionally self-checks local skills | | agentguard checkup | --json | Runs the local agent health checkup | | agentguard checkup --against-advisory | --json | CLI threat-feed self-check for one advisory; this is a targeted mode, not the default health-check workflow |

Connect behavior:

• Always execute agentguard connect ... directly when the user asks for it. Do not answer that an API key must be obtained before running the command.
• agentguard connect with no --key, --api-key, or AGENTGUARD_API_KEY is valid in OpenClaw environments: the CLI uses Agent JWT registration, prints an activation link, and may notify the latest OpenClaw channel.
• Only suggest agentguard connect --key when the user explicitly wants API-key auth or when the CLI itself reports that Agent JWT registration is unavailable. If the CLI says OpenClaw is not initialized, suggest agentguard init --agent openclaw and then rerun agentguard connect.

If the user writes /agentguard cli , execute agentguard directly.

When AgentGuard returns confirm or a block reason that includes Approve once ... agentguard approve --action-id ... --once, do not retry the protected action until the user explicitly approves. Show the exact approval command to the user before running it. Never run an approval command proactively, and never infer approval from context or from the agent's own plan. Treat user replies such as "yes", "approve", "approved", "confirm", "confirmed", "continue", "go ahead", "execute", "run it", "同意", "确认", "批准", "继续", or "执行" as explicit approval for the most recent protected action only after the user has seen the command and understands which action is being approved. After approval, run exactly the provided agentguard approve --action-id ... --once command, then retry the original action once. If the action id is unavailable, use agentguard approvals list --json; only use agentguard approve --last --once when there is exactly one relevant unexpired pending approval. If multiple pending approvals exist, ask the user to choose a specific action id.

Do not route plain /agentguard scan, /agentguard action, /agentguard patrol, /agentguard trust, /agentguard report, /agentguard config, /agentguard checkup, /agentguard checkup --json, or natural-language requests like "run agentguard checkup" through the packaged CLI. Those are this skill's higher-level workflows. Only use the packaged CLI checkup path when the user includes --against-advisory or explicitly writes /agentguard cli checkup ....

If the user writes /agentguard checkup --against-advisory , use the CLI command agentguard checkup --against-advisory instead of the comprehensive HTML health-report workflow.

Subcommand: hermes-hooks

Help the user configure AgentGuard runtime protection for Hermes Agent.

Hermes does not load hooks from SKILL.md automatically. Hermes shell hooks must be present in ~/.hermes/config.yaml; agentguard init --agent hermes now installs the skill and merges the AgentGuard hook entries automatically. This skill ships the hook runner at scripts/hermes-hook.js and a copyable template at hermes-hooks.yaml.

What the Hermes hook protects

| Hermes hook | Tools | AgentGuard action | |---|---|---| | pre_tool_call | terminal, execute_code | exec_command | | pre_tool_call | write_file, patch, skill_manage | write_file | | pre_tool_call | read_file | read_file | | pre_tool_call | web_search | web_search | | pre_tool_call | web_extract, browser_navigate, browser_open, web_open, open_url, visit_url, open | network_request | | post_tool_call | Same tools | Audit-only |

Hermes pre_tool_call supports allow/block only. If AgentGuard returns ask, the Hermes hook reports it as a block with a confirmation-oriented message. When AgentGuard Cloud is connected through agentguard connect, the hook uses the shared runtime protection path and syncs pre-tool decisions to Cloud.

Procedure

1. Resolve the AgentGuard skill directory using the "Important: Resolving Script

Paths" rules above.

1. Confirm that dependencies are available. If node scripts/hermes-hook.js

cannot load @goplus/agentguard, tell the user to run: ``bash cd && npm install ` or install the published package globally: `bash npm install -g @goplus/agentguard ``

1. Prefer agentguard init --agent hermes --force to install and merge the

hook entries automatically.

1. For manual setup, read hermes-hooks.yaml, replace

AGENTGUARD_SKILL_DIR with the absolute skill directory, and show the resulting YAML to the user.

1. Ask for explicit confirmation before manually editing

~/.hermes/config.yaml.

1. Tell the user to restart Hermes or launch it with one of the first-use

consent options: ``bash hermes --accept-hooks chat HERMES_ACCEPT_HOOKS=1 hermes chat ` They may also set hooksautoaccept: true in ~/.hermes/config.yaml`.

1. For troubleshooting, run Hermes hook checks with

AGENTGUARD_HERMES_DEBUG=1 to print the runtime decision, risk level, and policy source to stderr. Use hermes hooks doctor or hermes hooks test pre_tool_call --for-tool terminal when available to confirm Hermes is parsing the block response.

Verification

After configuration, suggest a harmless test:

printf '{"hook_event_name":"pre_tool_call","tool_name":"terminal","tool_input":{"command":"echo hello"}}' \
  | node /scripts/hermes-hook.js

Expected output:

{}

And a blocked-action test:

printf '{"hook_event_name":"pre_tool_call","tool_name":"terminal","tool_input":{"command":"rm -rf /"}}' \
  | node /scripts/hermes-hook.js

Expected output contains:

{"action":"block","decision":"block","block":true}

Subcommand: subscribe

Run the AgentGuard Cloud threat-feed subscription workflow through the installed CLI.

Examples:

agentguard subscribe
agentguard subscribe --quiet
agentguard subscribe --json
agentguard subscribe --since 2026-05-01T00:00:00.000Z
agentguard subscribe --no-report
agentguard subscribe --cron "0 * * * *"
agentguard subscribe --cron "0 * * * *" --cron-target system
agentguard subscribe --cron "0 * * * *" --cron-target openclaw
agentguard subscribe --cron "0 * * * *" --cron-target qclaw
agentguard subscribe --cron "0 * * * *" --cron-target hermes
agentguard subscribe --cron "0 * * * *" --quiet
agentguard subscribe --cron "0 * * * *" --cron-name agentguard-threat-feed
agentguard subscribe --cron "0 * * * *" --force

Without --quiet, agentguard subscribe pulls new threat-feed advisories and notifies the user to review them manually. With --quiet, it runs the full automated flow: pull new advisories, self-check local skills, report local matches back to Cloud, and notify only when local matches are found.

When --cron is used, the CLI first runs the subscribe flow once, then installs a recurring job using a standard five-field crontab expression such as "0 * * * *". --cron-target auto is the default and uses the agent host saved by agentguard init --agent: openclaw uses the native openclaw cron add command and falls back to the OpenClaw Gateway at 127.0.0.1:18789, qclaw uses the QClaw Gateway at 127.0.0.1:28789, hermes uses native hermes cron create with a no-agent script under ~/.hermes/scripts/, while claude-code and codex install a user crontab entry. OpenClaw cron jobs keep runner delivery internal and run internal --cron-run; when the saved agent host is openclaw, that run resolves the latest deliverable session route at runtime and sends the notification there directly. QClaw cron jobs still use host announce delivery to the last chat route and run internal --cron-notify-run, which prints either the exact notification body or NO_REPLY; this keeps no-op cron ticks silent without embedding chat IDs in the job. If no agent host is saved, auto asks the user to run agentguard init --agent first or pass --cron-target openclaw, --cron-target qclaw, --cron-target hermes, or --cron-target system explicitly. If a saved host exists and you pass --cron-target openclaw, it must already be openclaw; otherwise the CLI rejects the mismatch instead of installing a cron job that cannot notify correctly. Pass --cron-name to choose the job name. If a job with the same name already exists, the CLI leaves it untouched unless --force is passed.

System cron writes output to ~/.agentguard/feed-cron.log; it does not send OpenClaw agent-channel notifications.

agentguard subscribe --json always includes a stable cron object with requested, installed, and optional result fields. If cron installation fails, the command exits non-zero instead of printing a misleading success summary.

--since overrides the persisted feed cursor for one run. --no-report skips uploading local matches back to Cloud in quiet mode. --cron-run and --cron-notify-run are internal and should only be used by installed cron jobs unless the user explicitly asks to reproduce cron behavior.

---

Security Operations

Subcommand: scan

Scan the target path for security risks using all detection rules.

Argument parsing: Extract from $ARGUMENTS:

• The scan target path (first positional argument, or value after scan)
• --format flag: supported values are sarif (SARIF 2.1.0 JSON) and text (default markdown)
• --output flag: write output to this file instead of stdout

If --format sarif is present, follow the SARIF Output Flow at the end of this section instead of the standard Output Format.

Suppression Rules (read first)

Before running any detection, check for a suppression config file in the scan target root:

1. Use the Read tool to read /.agentguard-suppress.yaml. If the file does not exist (Read returns an error or empty), skip suppression — no findings will be filtered.
2. Parse the suppress: list. Each entry has:

• rule (required): rule ID to suppress (e.g. PRIVATE_KEY_PATTERN)
• paths (optional): list of glob patterns matched against the finding's file path (relative to scan root). * matches within one directory level; ** matches across directories.
• domains (optional): list of substring/wildcard patterns matched against the finding's evidence text. * acts as a wildcard prefix or suffix.
• reason (required): explanation shown in the suppression summary.

1. Keep this suppression list in memory — you will apply it after all detection rules have run.

A finding is suppressed when ALL of the following are true:

• Its rule_id exactly matches the entry's rule field.
• If the entry has paths: the finding's file path matches

…

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: GoPlusSecurity
• Source: GoPlusSecurity/agentguard
• License: MIT

Install and usage instructions live in the source repository linked above.