More screenshots
Works with the AI Agents you already use
voice agent guard
A deterministic safety framework for AI agents to block, confirm, or allow shell commands and tool calls.
$5
voice agent guard
Example session with this skill installed
Review the command 'rm -rf /' and explain how the guard should handle it based on the guard rules.
- Read your context and instructions
- Compiled the voice agent guard
The provided command rm -rf / is a critical system operation that violates multiple safety boundaries. According to the guard rules, this call must be intercepted at the PreToolUse stage and stopped immediately.
Guard Analysis: rm -rf /
| Check | Result | Reason |
|---|---|---|
| Command | rm -rf / | Recursive force-delete of the root directory. |
| Tier | BLOCK | Classified under "deletes system folders or a whole drive". |
| Response | deny | Immediate termination of the tool call. |
| Logic Path | Principle 1 & 2 | Block beats confirm. Root access is strictly prohibited. |
The guard identifies the rm verb with recursive and force flags targeting the root /. Any command attempting to delete system folders, the home directory, or a volume root is hard-blocked without a confirmation prompt. The agent should be informed that the action was denied and instructed not to attempt a workaround.
Hook Output
{
"hookSpecificOutput": {
"hookEventName": "PreToolUse",
"permissionDecision": "deny",
"permissionDecisionReason": "deletes system folders or a whole drive"
}
}
Next steps
- Ensure your
PreToolUsehook handles both the aliasdeland the verbrm. - Add a unit test specifically for root-relative paths in the block list.
- Configure the agent to report the specific
permissionDecisionReasonto the user via voice.
Connects securely to your tools. The creator never sees your data.
About this skill
Voice-driven AI agents can mishear commands or act on instructions hidden in web pages, leading to deleted files or leaked secrets. This skill provides a tested rulebook for building a PreToolUse permission layer that decides which commands run, which ask first, and which never run.
What it does
- Three tiers sort every tool call into Allow, Confirm or Block. Block beats Confirm, Confirm beats Allow, and anything unknown asks first.
- Wrapper unwrapping checks commands hidden inside
powershell -Commandorbash -c, and checks every part of a chained or piped command. - Untrusted content treats text from web pages, files, emails and the clipboard as data, and asks before any change once the agent has read it.
- Secret protection blocks reads of browser profiles and saved passwords, SSH keys, AWS and cloud credentials, keychains and system folders.
- Permission modes offer five levels, from Manual (every change asks) to Bypass (nothing asks, but the Block list still blocks).
- Browser rules cover clicks that submit, buy, post or delete, plus password and card fields.
How it works
- Intercept the tool call as JSON from a
PreToolUsehook, before it runs. - Split the command on pipes,
;,&&, newlines, brackets and backticks, and unwrap shell wrappers. - Check every part against the Block, Confirm and Allow lists. One risky part decides for the whole command.
- Answer the agent with
allow,askordeny, plus a plain reason the user sees. No answer within 30 seconds means no.
Frameworks & tools
Written for Claude Code and Codex hooks, with the exact JSON output format included. The rules carry over to any agent that runs shell commands. Covers bash, zsh and PowerShell on Windows, macOS and Linux. Drawn from the guard in Nova, an open-source voice assistant, where these rules are covered by tests.
Why this beats prompting it yourself
A system prompt that says "be safe" can be talked around, and the model can mishear or misjudge. This skill moves the decision into code that checks the real command, not the model's description of it. It also tells you to make that code fail closed: if the check crashes or nobody answers, the action is denied.
Use cases
- Voice interfaces for hands-free coding, where nobody is watching the terminal.
- Browser agents that must not click Delete, Buy or Send without asking.
- DevOps guardrails that stop an agent from pushing to git, deploying or changing infrastructure without asking.
Known limitations
This is a rulebook, not a ready-made hook: you build the parsing and register the hook yourself. Pattern rules can be bypassed by unusual spellings, so test them and expect to add rules over time. Database commands (for example DROP TABLE) are not covered. The browser rules assume your browser tool can inspect the real page element.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Security.
Only covers standard shells (bash, powershell) and common CLIs. Implementation of the parsing logic and hook registration is required by the developer.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 5 days ago
- Software
- jest
Node.js
Git