> ## Documentation Index
> Fetch the complete documentation index at: https://docs.agentopfor.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Browser extension

> Red-team any chat UI from your browser — no code required.

The browser extension is opfor's no-code path. Install it, open any chat interface, click the icon, pick a suite, and watch it run. It's built for everyone who doesn't open a terminal — PMs, QA, designers, and security analysts.

Everything runs **client-side in your browser** using your own LLM API keys. Nothing is sent to an opfor backend.

<Note>The extension covers **agent / chatbot** red-teaming. For MCP servers, use the [CLI](/cli/overview) or [MCP server](/mcp-server/overview).</Note>

## What it does

* Auto-detects the chat widget on any page (custom UIs, Intercom, Zendesk, Drift, Salesforce, etc.)
* Types attack prompts into the chat as if you were typing them, and watches the responses
* Judges each response with an LLM and scores it pass / fail
* Generates a self-contained HTML report you can download and share

It uses the same agent evaluator catalog as the CLI.

## Install

<Steps>
  <Step title="Add to your browser">
    Open the [OPFOR listing on the Chrome Web Store](https://chromewebstore.google.com/) and click **Add to Chrome**. Works in any Chromium browser (Chrome, Edge, Brave, Arc).
  </Step>

  <Step title="Pin the icon">
    Pin the OPFOR icon to your toolbar so it's one click away on any page.
  </Step>
</Steps>

<Accordion title="Development install (from source)">
  For contributors or testing unreleased changes:

  ```bash theme={null}
  git clone https://github.com/KeyValueSoftwareSystems/agent-opfor.git
  cd agent-opfor
  npm install
  npm run build:catalog --workspace=@keyvaluesystems/agent-opfor-extension
  ```

  Then open `chrome://extensions`, enable **Developer mode**, click **Load unpacked**, and select the `runners/extension/` folder.
</Accordion>

## Configure your LLM

The extension uses a single LLM configuration for attack generation, judging, and HTML parsing. It's stored in `chrome.storage.local` on your machine.

| Setting      | Description                                                                     |
| ------------ | ------------------------------------------------------------------------------- |
| **Provider** | OpenAI, Anthropic, Google, Groq, DeepSeek, Azure, or Custom (OpenAI-compatible) |
| **Model**    | Model name — pick a common one or type your own                                 |
| **API Key**  | Your key for the selected provider                                              |

<Tip>Use a capable model like GPT-4o or Claude Sonnet. Verdict quality depends on the model's reasoning ability.</Tip>

The gear icon exposes advanced settings: turns per attack (1–20, default 10), wait-after-send, message length limit, a manual agent description (fallback when auto-detect fails), a custom attack objective, business context, and an extra judge hint.

## Run a scan

<Steps>
  <Step title="Open the target">
    Open the chat interface you want to test in a browser tab.
  </Step>

  <Step title="Launch opfor">
    Click the OPFOR icon, then pick a **Security Suite** (e.g. OWASP LLM Top 10) or choose **Custom Evaluators** to select individual tests.
  </Step>

  <Step title="Set your LLM">
    Confirm provider, model, and API key, then click **Execute**.
  </Step>

  <Step title="Watch and download">
    The attacker types into the chat, the target replies, and live bubbles show the conversation. When it finishes, view the verdict and **Download report**.
  </Step>
</Steps>

The extension runs up to 20 turns per evaluator (default 10) and stops an evaluator early once the judge returns a definitive verdict.

<Card title="What it tests" icon="shield-halved" href="/evaluators/reference">
  Same agent catalog as the CLI — full evaluator and suite reference.
</Card>

## Limitations

* **Chat-widget targets only.** It needs a detectable chat input + response area in the DOM. For standalone API endpoints or non-chat agents, use the [CLI](/cli/overview).
* **No MCP or live tool-call evaluators.** The judge sees the transcript, not real tool side-effects. For MCP servers use the CLI's `mcp` mode or the [MCP server](/mcp-server/overview).
* **One agent per run** — no inter-agent tests.
* **Vendor closed shadow-DOM widgets** (Salesforce, some Intercom builds) may need a vendor-specific fallback. Open an issue with a sample URL if auto-detect fails.
* **No pause / resume across sessions** — run state lives in `chrome.storage.local`; uninstalling wipes it.

## Privacy

* API keys live in `chrome.storage.local` and never leave your browser. There is no opfor-hosted backend.
* Attack prompts and target responses go only to the LLM providers you configure — the same data path as running the CLI on your own machine.
* The extension does not phone home.