This guide takes you from zero to a working, sanitized AI call. You'll create a project, apply a starter policy, connect a model, and watch the Gateway redact sensitive data before it leaves. Everything on this page configures one thing: the AI data-privacy and governance gateway that stands between your app and the model.
Before you start
You'll need an intSignal account with the AI Gateway enabled and permission to create projects (an admin or developer role). Sign in at the Network Portal.
1. Create a project
Open the AI Gateway
In the Portal, go to AI Gateway → Projects and choose New project. A project is a boundary for policy, model routing, keys, and logs — typically one per app or per client.
Pick a starter policy
Choose a built-in policy such as Standard PII, HIPAA, or PCI, or start from Automatic and let the Gateway propose rules from your data. You can tune it later in Policies.
Connect a model
Add a model provider — OpenAI (GPT), Anthropic (Claude), Azure OpenAI, or a private endpoint — and paste that provider's key. The Gateway stores it as a secret and calls the model on your behalf. See Models & agents.
Create a Gateway key
Under Settings → Keys, create a key scoped to this project. This is the key your app uses to call the Gateway — not the model provider's key.
2. Send your first sanitized prompt
Call the Gateway exactly where you'd normally call a model. It sanitizes the input, forwards it, checks the reply, and returns it.
import { Gateway } from "@intsignal/ai-gateway";
const gateway = new Gateway({ apiKey: process.env.INTSIGNAL_GATEWAY_KEY });
const result = await gateway.chat({
project: "support-assistant",
model: "gpt-4o", // or "claude-sonnet", "azure:gpt-4o", "private:llama"
messages: [
{ role: "user", content: "Refund Jordan Rivera, SSN 123-45-6789, card 4485 1234 5678 9012." },
],
});
console.log(result.output); // the model's answer
console.log(result.redactions); // [{ type: "ssn", action: "redact" }, { type: "pan", action: "mask" }, ...]
The model receives a sanitized message — the SSN removed, the card masked — while your app still gets a useful answer. Nothing sensitive left your boundary.
3. Verify in the log
Open AI Gateway → Logs and select the request. You'll see the original input, the sanitized version that was sent, every entity that was redacted or blocked, and which model answered. This is the same evidence you can export for audits.
4. Put it in front of an existing assistant (no code)
If you don't control the app's source — for example a team using Microsoft Copilot or ChatGPT Enterprise — run the Gateway as a proxy and point the assistant's model endpoint at it instead. Same detection and policy, zero code changes. See Integrations.
Next steps
- Define exactly what to redact and block in Policies.
- Let the Gateway suggest rules with Automatic policy setting.
- Add document capture (scan-to-text, scan-to-form).
- Explore the full AI Gateway API.