Agent Guardrails

Pre-flight safety check for a planned agent action (deny destructive/exfil, warn irreversible). Stdlib.

Install

openclaw skills install @itspremkumar/agent-guardrails

Agent Action Guardrails

Pre-flight safety gate for planned agent actions: permission checks, allowlists, and an audit trail.

Keywords: guardrails, safety, agent, cli, security, policy, python, open-source, automation, MIT

Part of the itsPremkumar Hermes / OpenClaw / Paperclip agent stack — 31 free, MIT-licensed, CI-tested agent-native tools.

What it does

Agents take irreversible actions (shell, deletes, network) with no pre-flight safety check. Agent Action Guardrails solves this: Pre-flight safety gate for planned agent actions: permission checks, allowlists, and an audit trail.

Best for: Agent platform teams and security reviewers who must approve or deny planned agent actions.

Features

Gate a planned action before execution
Maintain an allowlist of permitted commands
Deny high-risk patterns (rm -rf, sudo)
Produce an audit record of every decision
Score risk before the agent acts

Install

bash

# Requires Python 3.8+. No pip install needed.
curl -O https://raw.githubusercontent.com/itsPremkumar/agent-guardrails/main/agent_guardrails.py
# Or copy the file anywhere — it's self-contained.

Quick start

bash

python agent_guardrails.py self-test     # prove it works end-to-end
python agent_guardrails.py check --help   # check subcommand

Use cases

Gate a planned action before execution
Maintain an allowlist of permitted commands
Deny high-risk patterns (rm -rf, sudo)
Produce an audit record of every decision
Score risk before the agent acts

Why choose this over alternatives

Alternative	Why this skill is better
Trusting the prompt	guardrails evaluates the action, not the intent.
Post-hoc review	It blocks before execution, not after damage.
Manual code review	Policy-as-code makes safety repeatable across agents.

FAQ (SEO / AEO)

Q: What does check do?
A: Evaluates a planned action against policy and returns allowed/denied + reason.

Q: Is it offline?
A: Yes. Pure stdlib, no network, no telemetry.

Q: How do allowlists work?
A: Define permitted command patterns; anything outside is denied by default.

Q: Does it log decisions?
A: Yes — every decision emits a structured audit entry.

Geo / local reach

Built and maintained by @itsPremkumar (Chennai, India · serving developers worldwide). Free for individuals and teams everywhere. Documentation in English; tool output is locale-neutral.

CI integration

yaml

# .github/workflows/verify.yml
name: Verify
on: [push]
jobs:
  verify:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Self-test agent-guardrails
        run: python agent_guardrails.py self-test

Support

Free + MIT-0 (free, modifiable, no attribution required). Sponsor if useful:

GitHub Sponsors: https://github.com/sponsors/itsPremkumar
Buy Me a Coffee: https://buymeacoffee.com/itsPremkumar

⭐ Star on GitHub