Live demo · In testing

Sentragate

A Zero Trust gateway for AI workloads. Every request to a model is verified, checked against policy, rate limited and inspected, with the decision logged. Try to get past it.

1. Sign in as

61/2000

Prompts are inspected and discarded, never stored. The model replies are canned: no real LLM is called.

Gateway decision

  1. Identity
  2. Policy
  3. Rate limit
  4. Guardrails
  5. Model

Pick an identity, a model and a prompt, then send it. Each request is verified, checked against policy, rate limited and inspected before anything reaches a model.

Your audit trail

Every decision the gateway makes for you is logged here. You only ever see your own entries, and they expire after an hour.

How it works

01

Identity

Each request carries a short-lived signed token. In production this is a Microsoft Entra ID access token; here the site signs a 2-minute demo token for the persona you pick.

02

Policy

Conditional Access-style rules weigh group membership, device compliance and sign-in risk against the model requested. Deny rules always win, and anything unmatched is denied.

03

Rate limit

Limits are tied to the verified identity, not a shared API key.

04

Guardrails

Prompts are scanned for PII, leaked secrets and prompt-injection phrasing before anything leaves the gateway.

Source and threat model on GitHub.