Open source (MIT)

Stop your AI agent before it breaks prod.

Open source (MIT). Built from the guards we run on our own systems every day. Pre-execution command interception that blocks destructive ops, secrets leaks, and PII exposure at the shell level. Try Pro free for 14 days, then 29 EUR/month, cancel anytime.

Install GuardRailTry Pro FreeView source
Guards

14 core guards. Each one blocks one specific mistake.

Each guard solves one specific problem. No ML, no prompt engineering. Pattern matching at the command level. Six examples:

Blocker
Destructive Paths

Blocks rm -rf on critical paths and root-level deletions.

destructive_path_guard
Blocker
Mass Update

Blocks SQL UPDATE and DELETE without a WHERE clause.

mass_update_guard
Blocker
Protected Branches

Blocks direct pushes to protected branches and destructive git operations.

main_push_guard
Blocker
Secret Exfiltration

Blocks commands that send secrets to known exfiltration domains and secrets written into files.

basic_secret_detector, edit_secret_guard
Blocker
Services and Firewall

Blocks stop, disable or kill on critical services and firewall flushes.

service_protection_guard, firewall_flush_guard
Detector
Injection and Env Dumps

Detects prompt-injection patterns and environment dumps in command output.

basic_injection_scanner, env_dump_detector
How it works

Three steps.

GuardRail intercepts at the shell level. No SDK, no wrapper, no proxy. Pure bash.

Intercept

Hook into your AI agent's shell dispatcher. Every command passes through GuardRail before execution.

Evaluate

Pattern matching against the installed guard rules. No ML and no API calls in the core guards. Deterministic decisions. Blockers halt, detectors warn.

Act

Blocked commands return non-zero with a clear message. Audit trail logs every decision. Override with explicit confirmation when needed.

Get started

Two commands.

Step 1: Install 14 core guards

npx guardrail-agent init

Step 2 (optional): Unlock 48 Pro guards, 14 days free

Start 14-day Pro trial

Card required, nothing is charged for 14 days. Then 29 EUR/month, cancel anytime. Your license key arrives by email: guardrail upgrade <key>

Requires bash 4+, jq. Linux and macOS. Works with Claude Code, Cursor, Aider, and any shell-based agent.

Pricing

Open core. Pay for what you need.

The 14 core guards are MIT licensed. Pro adds 48 guards, the EU AI Act report and support.

Community
0 EUR / open source

For individual developers.

  • 14 core guards
  • Secret and env-dump protection
  • Protected-branch guard
  • Audit trail
  • Pentest command
  • Community support (GitHub)
Get Started
14 days free
Pro
29 EUR / month

For teams running agents in production.

  • 48 Pro guards from 14 months of real incidents
  • Everything in Community
  • PII scanning of agent output
  • EU AI Act compliance report (article mapping)
  • Multi-step attack detection
  • Email support
Start 14-day free trial

Then 29 EUR/month, cancel anytime. Purchase and invoice via Stripe; statutory rights apply. Buy without trial

Custom guards or a setup for your team? Book a call or write to frederik@frederikvonderheyden.de.
Template Pack

AI Governance Starter Kit

20 guards, 4 skill runbooks, dispatcher templates and a CLAUDE.md template from our own system. Sales are paused while we revise the kit.

Starter Kit: back soon

Includes: 20 guards, 3 dispatchers, 4 skills, CLAUDE.md template, settings.json example, customization guide.

Go Deeper

KI-Governance Masterclass

30 lessons from 14 months of running autonomous AI agents in production. Learn the complete GRIP framework, build your own guard system, and set up EU AI Act compliance.

Masterclass: back soon Course outline

6 modules, 30 lessons. Guards, Reinforcement, Intelligence, Protocol, Compliance, Autonomous Ops.

The Book

Runs Without Me. You Can Too.

GuardRail was built over 14 months as a solo founder. The book shows the complete system: GRIP framework, Crystallization Loop, the guard system and the architecture behind it.

Buch kaufen (DE)Buy Book (EN)