Massachusetts employer? AI training for your team can be up to 100% reimbursed. See the MA grant
← All plugins

Agentic Guardrails coding plugin

Protect your files from risky AI changes with Agentic Guardrails

Keep AI coding changes recoverable. Agentic Guardrails checks risky file and command actions in supported coding sessions and steers destructive requests toward safer, reversible paths.

The problem

A coding agent asked to tidy a project may remove a folder or overwrite a file you meant to keep. Those changes can be hard to undo after the command runs. You need the agent to handle risky actions with care while the work is happening, with a way back when an operation should be reversed.

The solution

We built Agentic Guardrails to check supported file and command actions as a coding agent works. Its trusted hooks can stop a direct deletion, guide the agent toward a reversible archive, and keep operations within the task's stated bounds. You remain able to review the decision and recovery path.

How it works

  1. Set the task bounds

    Tell the agent which files it may change and which work it must preserve. Guardrails uses the stated task intent when checking supported actions.

  2. Check proposed actions

    Active hooks inspect supported file and command actions before they run. A risky operation can be blocked or sent back for review.

  3. Use a reversible path

    For a direct deletion, Guardrails can deny the request and point the agent toward agw archive instead of removing the folder outright.

  4. Review and recover

    Review what the agent changed and the archive it created. Use the documented recovery path if you need to restore an archived item.

Use Agentic Guardrails in your app

New to plugins? Read the beginner guide. Choose your desktop coding app below, install and trust the hooks, then follow the product's disposable-folder test. Inspect the host's actual hook decision and any archive guidance before using Guardrails on a real project.

Add the plugin in Claude desktop, then use it in the desktop Code tab with active hooks.

Set it up

  1. Download Claude desktop if needed and sign in.

    Download Claude ↗
  2. Open Customize > Plugins > Add > Add marketplace > Add from a repository. Enter this address.

    ProfSynapse/synaptic-labs-plugins
  3. In Discover, find Agentic Guardrails and choose Add.

  4. If Python 3.9 or newer is missing, use the product setup guide or ask your coding assistant to help install it for your operating system.

    Guardrails setup guide ↗
  5. Switch to the Code tab and start a new session in a disposable folder. Review and trust the plugin hooks in the host prompt.

  6. Ask the coding agent to run agw doctor --json for health, then follow the verification guide to test a deletion. Inspect the actual hook decision and any agw archive guidance.

First use

Check the hook on a test action

Open a Claude desktop Code, ChatGPT desktop, Claude Code, or Codex session with trusted Guardrails hooks in a disposable folder.

Create ./scratch-dir, then ask the agent to delete it using the official host verification guide. Do not use a real project folder.

What to expect

Inspect the hook decision and any guidance toward agw archive; a doctor health result alone does not prove interception.

Your review

Verify the actual host response and recovery path before trusting Guardrails on real files.

Frequently asked questions

Can I set it up from a desktop app?

Yes. Claude desktop has a Plugins marketplace and a Code tab; ChatGPT desktop also has a Plugins marketplace. Install Guardrails, approve its hooks, and start a new coding task. Claude Code and local Codex have command-line setup paths too.

Does installing it protect every AI chat?

No. Its checks depend on trusted, active hooks in a supported coding session. An ordinary chat without those hooks is not protected by the Guardrails file-action checks.

How do I know the hooks are working?

Use a disposable folder and the product verification guide to request a test deletion. Inspect the host's actual hook decision and any agw archive guidance. agw doctor --json reports setup health, but is not proof that an action was intercepted.

Can I recover a file action?

Guardrails routes supported destructive operations toward a reversible archive, with recovery commands in its guide. Check the archive entry and recovery path for your specific action before depending on it.

Does it guarantee that no file can be lost?

No. Hook support, trust, and the action path matter. Keep normal backups and verify the plugin in your host. Windows is the stable platform; macOS and Linux support is preview.