What is it?

Endpoint Agent Defender connects popular AI agent installed on the user’s endpoint device (Cursor, Codex, GitHub Copilot, OpenClaw, OpenCode) to the Pragatix Firewall on the user’s PC.

When someone uses an AI agent to chat, run tools, or start a session, the defender can:

• Record activity in Pragatix for compliance and visibility

• Apply policy so risky prompts and tool use can be blocked before it runs

Setup is a short Windows installer plus a one-time choice of which AI tools to protect.

Install here:

Why should customers care

• One control point across the AI tools developers already use

• Security and governance without banning AI outright

• Fast rollout - installation in seconds via a guided installer

• Works alongside existing Pragatix Firewall licensing and auditing

What you need before install

For staging use the following: ​https://staging-firewall.pragatix.ai/

The installer can install Python automatically if it isn’t already on the machine (required for hooks to run).

How installation works

1. Download and run Install-EndpointAgentDefender.ps1(attached above).

2. The installer prepares Python if needed.

3. It downloads or reuses the defender package.

4. On first run, it asks for Firewall URL and API key.

5. The user picks which AI agents to set up (Cursor, Codex, GitHub Copilot, OpenCode, OpenClaw).

6. Hook files are copied into each product’s profile automatically.

7. User restarts the AI apps they enabled.

Agent Presentation Guide & Notes

OpenCode

1. Download bun - run this in terminal:

   powershell -c "irm bun.sh/install.ps1 | iex"
   

2. Install open code - run this in a new terminal window:

   bun add -g opencode-ai
   

3. Run the Endpoint Defender installer - download here:

4. Follow install instructions:

   • For the firewall URL, enter:
      https://bgd15-pragatixapi.agatdemo.com

   • For the API key, copy from AI Guardian → AI Gateway

5. Run open code web UI - run in terminal:

   opencode web
   

6. Set block policy: Go to AI Security Suite → Guardian Agent → Agent Policies, and set the following BLOCK policy: Opencode → Opencode Built-In Tools → Action Intent → Delete Operations

7. Ask the agent to delete a file

How to confirm hooks are working

Use these simple checks in a pilot. Add your own screenshots where noted.

Quick test (any agent)

1. Open the AI tool and ask a question.

2. Open the local activity log (see Help below) - there should be a new entry with current date and time.

3. In Pragatix Firewall, confirm the activity shows up in prompt/agent auditing.

Cursor

• Go to Settings → Hooks

• Hooks and their logs should appear:

Codex

• Codex indicates the hooks that were fired in the chat window itself:

• Clicking on the hook icon opens a window with details on the hooks fired:

GitHub Copilot (in VS Code)

• From the chat window: click Open Customizations (gear icon) → Hooks

• Alternatively, you can write “>hooks” in the searchbar and press Enter

• Number indicators will show next to found hook types:

OpenClaw

• Plugin copied under .openclaw\extensions.

• User must turn the plugin on in OpenClaw settings.

• Run a tool, then check the openclaw log folder.

[Screenshot: plugin enabled]
[Screenshot: log updated]

OpenCode

• Same idea: use the product, then confirm the opencode log folder has new entries (when hooks are included in the bundle).

[Screenshot: OpenCode + log]

Help

Where to find logs

On the PC, open the defender install folder (usually C:\.agat\PragatixAgentDefender), then navigate to:

agents\<product name>\logs\policy-hook.log

Examples:

The log file policy-hook.log is the main place to see whether hooks ran, what was sent to the Firewall, and whether the action was allowed or denied.

The log is created after the first real use of that agent. If the file is empty or missing, hooks may not be installed, the app wasn’t restarted, or Python wasn’t available.

Common issues (plain language)

Troubleshooting - What to send support

1. Which AI product (Cursor, Codex, etc.)

2. Whether the installer reported success for that product

3. A recent snippet from policy-hook.log (verify no sensitive information is included).