What is Sentinel?

SUPERWISE® Sentinel is an AI firewall: a single control plane to secure and standardize decentralized AI operations. It sits in front of your LLM traffic as a lightweight proxy, enforcing guardrails at the edge and giving you full visibility, without your prompts or responses ever leaving your environment.

SUPERWISE Sentinel — the AI firewall in front of your organization's LLM traffic

Prompts and responses stay between you and the model provider you chose — only operational metadata reaches SUPERWISE. See Data privacy.

How does a request flow through Sentinel?

flowchart LR
  tools["Your AI tools<br/>Claude Code, your own apps,<br/>ChatGPT and Claude in the browser"]
  gateway["Sentinel gateway<br/>on your PC or in your network<br/>checks the prompt and<br/>replaces sensitive details"]
  provider["The model provider you chose<br/>OpenAI, Anthropic, or Gemini"]
  log["Live activity<br/>the request's metadata is recorded"]
  tools -- "prompt" --> gateway
  gateway -- "cleaned prompt" --> provider
  gateway -. "what was checked and replaced" .-> log
  1. Your AI tools send the prompt to the Sentinel gateway. The gateway runs on your PC or in your network, and the tools reach it instead of the provider. There is no SDK, wrapper, or code change: the Windows app and the CLI point supported tools at the gateway, and your own apps reach it through the provider's base URL.
  2. The gateway checks the prompt against your guardrails. Secrets and personal data are replaced with a placeholder, {{REDACTED}} by default, and a request that should not be sent at all is blocked.
  3. The cleaned prompt goes to the model provider you chose, such as OpenAI, Anthropic, or Gemini, as it would without Sentinel. The response comes back to your tool unchanged.
  4. The request's metadata is recorded. Which guardrails ran and what they did appear in Live activity. SUPERWISE receives this operational metadata, never the prompt or the response.

Nothing changes in how your team works.

Where does Sentinel run?

Prompts and responses never pass through SUPERWISE's cloud. Each gateway reports only operational metadata (request counts, which guardrail fired, and latency) to the SUPERWISE control plane, and pulls its policy from it. See Data privacy.

Which AI tools does Sentinel cover?

Tools running on the machine are covered by the Sentinel for Windows app or the Sentinel CLI; coverage depends on the tool and the OS, as below. Chat sites are covered through the browser extension, which is in preview.

On the machine

Terminal agents, desktop apps, and the apps you build.

ToolWindowsMac & Linux
Claude CodeSupportedSupported
Codex CLISupported*Supported*
Antigravity (agy) CLI, Google's coding agentSupported*Supported*
Your own apps and agents (OpenClaw, LangChain, or any framework)SupportedSupported
Claude DesktopSupportedSupported
ChatGPT / Codex desktop appComing soonSupported
Antigravity / Gemini desktop appComing soonComing soon

*The Codex and Antigravity CLIs are covered when they authenticate with an API key, not when signed in with a ChatGPT or Google account. On Mac and Linux, the Sentinel CLI configures Codex and Claude Desktop for you.

In the browser

The browser extension is in preview, so all browser coverage here is preview. It works in Chrome and Edge, on Windows only for now, alongside the Sentinel for Windows app.

Chat siteStatus
Claude (claude.ai)Supported (preview)
ChatGPT (chatgpt.com)Supported (preview)
Gemini, Copilot, Grok, Perplexity, DeepSeekIn development

What does Sentinel not cover?

Sentinel replaces the identifiers its guardrails recognize, listed in PII redaction and Secret & credential detection. It does not cover names, street addresses, other dates, or unlabeled record numbers, so names and most dates still reach the model. It acts only on the tools marked Supported or Preview above.

Sentinel works by base-URL redirection: a tool is covered when its provider base URL points at the gateway. It does not sit in front of all network traffic, and a tool that calls the provider directly is not checked. On the requests it receives, the gateway checks the text-generation routes, OpenAI v1/chat/completions, Anthropic v1/messages, and Gemini generateContent, and passes every other route through unchecked. Guardrails run on the prompt; the model's response comes back unmodified. See Architecture.

See also

  • Get started: install Sentinel and run your first check.
  • Architecture: how the control plane and gateways fit together.
  • Data privacy: what stays on your side and what reaches SUPERWISE.

Did this page help you?