A single host with Docker Compose
The reference installation: one compiled image next to your PostgreSQL, behind your reverse proxy.
Install guideInstall InfraGuard where your servers, databases and models already are. It controls internal access for agents and people, keeps its record on your database, and needs nothing from SoftQuantus to run.

Where it runs
The reference installation: one compiled image next to your PostgreSQL, behind your reverse proxy.
Install guideThe chart runs the compiled image, reads every secret as a file and runs each migration as its own step. It creates no Secret.
Install on KubernetesVerify and load the offline bundle with no network and no registry. The licence is checked offline too.
Release verificationBefore anything runs
The release key comes from the documentation site, never from the bundle. The verification script stops at the first failure and prints REFUSED; a bundle you cannot verify is not a release.
What you provide
Network paths
| From | To | How |
|---|---|---|
| People and tools | InfraGuard | HTTPS through your proxy |
| Collectors | InfraGuard | Outbound HTTPS only; collectors have no inbound port |
| InfraGuard | PostgreSQL, identity keys, OpenSearch, models | Only the services you configure |
| InfraGuard | SoftQuantus | Nothing. No call home, no online licence check |
Your vault agent or the Secrets Store CSI driver writes each secret to a file; InfraGuard reads the file and picks up rotations.
Read moreReference NetworkPolicy, Docker internal network and nftables rules make the gateway the only path to your models.
Read moreThe program in the image is compiled. What stays readable is the database migrations, the console’s page, scripts and style sheet, and a version marker.
Read more
Request
POST /v1/llm/chat/completions
support-assistant vllm/qwen3-8b
Grant
allowThe model server's key stays inside InfraGuard.
Past the limit
429 · Retry-After 12 sThe 61st call in a minute gets a 429 with Retry-After. Revoke the grant and the next call is refused.
Tokens per day, last 7 days. Prompts are not stored.
Private AI bundle
The bundle puts LibreChat in front of an open-weight model from Hugging Face, Qwen3-8B served by vLLM, with InfraGuard as the gateway and Keycloak for sign-in if you have no identity provider yet. Prompts stay on your network, and every call passes the same grants and limits as your agents.
The files are delivered with the release; we bring the bundle up on your GPU host during the pilot.
Your estate
Servers, virtual machines, clusters, databases, directories and the logs around them, each reached the way the tile says, today.
Linux servers
Red Hat Enterprise Linux, Ubuntu, Debian, SUSE
SSH command profiles: typed parameters, never a free shell line
VMware vSphere
ESXi hosts and their guests
Linux guests through SSH command profiles
Microsoft Hyper-V
Windows Server virtualisation
Linux guests through SSH command profiles
Proxmox VE and KVM
Open-source hypervisors
Linux guests through SSH command profiles
People on servers
Engineers and administrators
Short-lived SSH certificates issued by InfraGuard
Kubernetes
OpenShift, Rancher and other distributions expose the same API
Pods, logs, events, restarts and discovery; tested on k3s
PostgreSQL
Operational databases
Read-only queries, writes under policy, temporary logins that expire
PostgreSQL in your cloud
Azure Database for PostgreSQL, Amazon RDS, Google Cloud SQL
The same PostgreSQL connector; not yet verified on these managed services
Microsoft Active Directory
Over LDAPS
People, groups and nested membership, synchronised on a schedule
Keycloak and OIDC
Single sign-on
Signed access tokens; tested end to end with Keycloak 26.4
SCIM 2.0 and SAML
Provisioning and federation
SCIM provisioning; SAML through an identity broker
OpenTelemetry
OTLP over HTTP
Logs and metrics in, each stored under its sender’s credential
OpenSearch
Version 3.9
A search mirror with one role per tenant
Your SIEM
Security operations
The audit chain by cursor, and signed webhook alerts
Hugging Face
Open-weight models from the Hub
Download once, serve on your GPUs, govern every call
vLLM
Model serving
The serving engine of the Private AI bundle
llama.cpp and Ollama
Local model servers
Any OpenAI-compatible or Ollama endpoint behind the gateway
LibreChat
Chat for employees
Private AI bundle: chat, model and sign-in on your machines
MCP and REST agents
Agent frameworks and assistants
One tool gateway over MCP streamable HTTP and REST
Docker Compose
A single host
The reference installation
Helm
Your Kubernetes platform
A chart with reference egress rules
Air-gapped networks
No internet at all
An offline bundle, verified locally, with an offline licence
Get started
InfraGuard is delivered as an offline bundle through a pilot. Verifying and loading it needs no network and no registry, and the running service makes no call to SoftQuantus.
cd infraguard-offline-<commit>
./verify-release.sh --key infraguard-release.pub --commit <commit>
# VERIFIED: infraguard-api-<commit>.tar.gz ...
docker load -i infraguard-api-<commit>.tar.gz
# Loaded image: infraguard-api:<commit>Requirements, configuration and the first steps are public. Read them before you talk to us, so the pilot starts on your questions.
Install guideTell us what your agents and people should reach. We agree a scope with you and send an offline bundle and, where the pilot needs one, a licence.
Request a pilotCheck the bundle with its own verification script before anything runs, then start InfraGuard with Docker Compose or the Helm chart, next to your PostgreSQL and identity provider.
Release verificationRegister one cluster, database or server, write the first policy and connect an agent over MCP. The tutorials walk through each step.
TutorialsTell us what your agents and people need to reach, and where InfraGuard would run. It takes about two minutes.
The film and pictures on this page are SoftQuantus’s own, painted from noise and line work; the clips are recordings of the InfraGuard tutorials. Product names belong to their owners and show where InfraGuard fits, not a partnership or an endorsement.