Skip to content
§00 / Sovereignty

Data never leaves your perimeter.

HA.Prizma deploys entirely inside your perimeter, from the gateway to the audit store. In your perimeter · under your control · with provable audit.

perimeter

Entirely in your perimeter

No prompt, document or key ever leaves your network.

  • Docker Compose or Kubernetes on your own hardware
  • No SaaS mode, by design
  • The control plane lives next to the models
air-gap

Runs in air-gap

Deploys in an isolated segment with no internet access.

  • Offline licenses signed with Ed25519
  • Verified by a built-in public key
  • Zero call-home at install and runtime
degrade

Degrade, never brick

An expired or missing license never drops protection.

  • The critical path does not depend on the license
  • Gateway, DLP, audit, kill-switch always work
  • The license gates only the optional
§01 / Quickstart

From distribution to a running gateway in four steps.

The delivery arrives in your perimeter from a private registry, one line of config routes traffic to the gateway, orchestration carries the same stack to a cluster. A starter stand comes up in ~2 min, and each tool connects in ~30 s.

  1. [01]

    Distribution into your perimeter

    Images and the chart arrive under license from a private registry straight into your perimeter. Nothing is pulled from public sources.

  2. [02]

    One line of config

    Any OpenAI- or Anthropic-compatible client goes through the gateway with no code changes: only the API address changes.

  3. [03]

    Orchestration

    The same stack deploys to a cluster with standard orchestration. Environment values override the registry, tags, secrets and ingress host.

  4. [04]

    Tools in ~30 s

    A developer connects a CLI or IDE tool themselves, self-serve, no ticket. DLP, budgets and audit apply at once.

§02 / HA and operations

Resilience, proven by outages.

A production topology with automatic failover, point-in-time recovery and observability out of the box.

0 data loss
automatic failover

proven by planned failure drills

built-in
metrics and alerts

Grafana dashboards and alerts

LB application data observability backups Traefik · HAProxy route to leader · TLS Backend · 1 gateway · DLP · audit Backend · 2 gateway · DLP · audit Postgres · master Patroni Postgres · replica hot standby etcd quorum 3/3 Redis Sentinel 3 · m/replica Prometheus autoscaler Grafana metrics · alerts WAL archive continuous PITR · backups to any point S3 · MinIO off-host streaming replication TRAFFIC A primary failure automatically moves load to the replica, with no data loss.

  • PostgreSQL · Patroni + etcd + HAProxy

    A primary failure never drops the control plane: etcd holds consensus by quorum, HAProxy routes to the current leader. Planned failure drills confirmed automatic failover with no data loss.

  • Redis · Sentinel

    Cache and pub/sub survive a node loss. A Sentinel quorum runs automatic failover; kill-switch and rate-limit keep propagating.

  • PITR · WAL archive

    Roll back to a point in time after an incident. A continuous WAL archive to off-host S3-compatible MinIO storage, recovery to any moment.

  • Prometheus autoscaler

    Throughput follows the load. On a live stand it scaled 1 → 5 → 1 workers by pipeline metrics, with no manual intervention.

§03 / Integrations

A tool connects in ~30 s.

One gateway for 12 CLI and IDE tools: one line of config or a self-serve connector. DLP, budgets and audit apply at once. 

curl https://gw.prizma/install/claude | sh

~30 s per tool · no SDK

output channels

VS Code extension

editor

DLP linter and clipboard-guard in the editor

MCP Gateway

mcp

agent tool-calls through the same DLP and audit

n8n node

workflow

LLM calls in no-code pipelines

Webhooks

alerts

DLP · kill-switch · budgets → Slack / Teams / Telegram

SIEM · CEF / syslog

siem

CEF and syslog export · e2e with Splunk

Self-serve connector

cli

curl-install the CLI, no ticket

§04 / Network enforcement

Every AI tool goes through the gateway.

Three layers of network enforcement leave no bypass to the cloud around DLP and audit, even for a tool security does not know about yet.

shadow-ai detection

Unauthenticated traffic is not blocked blindly: first it is visible.

Anonymous clients are identified by fingerprint (IP + UA). Policy: block, monitor or allow. A client can be turned into a real user in one action (provisioning). That is how shadow AI is found before the first incident.

  1. 01

    PAC file

    Proxy auto-config routes all traffic to api.openai.com / api.anthropic.com and other providers to the gateway.

  2. 02

    DNS override

    The internal resolver points provider domains at HA.Prizma, so bypass via a direct domain is closed.

  3. 03

    Firewall rules

    Egress to model APIs is allowed only from the gateway; direct exit from workstations is blocked at the perimeter.

Ready?

All your AI traffic under your control

A self-hosted LLM Firewall inside your perimeter. One base_url line, and no prompt leaves directly.

Request a demo How it works
sales@hackadvisor.io NDA on request · reply within one business day