StellarBase
Docs Deployment Air-gapped
Deployment

Air-gapped Deployment

Deploy the full platform inside a network with zero internet egress. Everything — LLMs, agents, workflows, admin UI — runs on your hardware. For classified workloads, defence, and regulated industries that cannot send data anywhere.

What air-gapped means here

“Air-gapped” is used broadly. We mean: no outbound internet connectivity at all from the StellarBase deployment. No phone-home, no telemetry, no license server check-in, no model updates over the network. If you disconnect the internet cable, the platform continues to function indefinitely.

What ships in an air-gapped install

Platform images

Signed container images for every service. Delivered as tarballs for import into your internal registry. No dependencies pulled from external registries at runtime.

Data packs

Content that a connected deployment would fetch from the internet — ships as signed offline packs:

  • LLM weights — GPT-OSS 120B, Devstral 2, Qwen 3.5 397B, or a subset. You choose which models to install.
  • Specialized model weights — Qwen3-Embedding, BGE-M3, DINOv3, GLiNER variants, StellarOCR components
  • Offline corpora — legal codes (Sbírka zákonů, EUR-Lex), scientific literature mirrors, regulatory databases. Optional — pick what you need.

Tooling

Helm charts, admin CLI, migration scripts, health-check tools — all shipped in the package.

Update model

Updates are delivered as signed data packs on physical media or via your organization’s controlled data-transfer process. A typical cadence:

  • Security patches — within 14 days of release, higher priority for critical
  • Minor releases — quarterly
  • Major releases — annual
  • LLM refresh — new model versions bundled with minor / major releases
  • Legal corpus refresh — quarterly (Sbírka zákonů, EUR-Lex)

Customers with strict change-management processes can skip any release — no forced upgrades.

What’s different from standard on-prem

Compared to internet-connected on-prem:

  • All LLM inference must be local — no routing to cloud providers
  • External APIs (OpenAI, Anthropic, etc.) simply aren’t available
  • Connectors to internal systems only (no SaaS connectors like Notion, Slack, Drive)
  • Updates are pull-based through your controlled transfer process
  • No telemetry, crash reports, or usage metrics transmitted

Capabilities that still work

Everything that doesn’t require an external service:

  • Full knowledge base, ingestion, search, retrieval
  • All agents, workflows, chat, collaboration
  • StellarOCR and all specialized models
  • Local LLMs (GPT-OSS, Devstral, Qwen, or subset)
  • StellarGate anonymization (useful even when you don’t call external LLMs — for logging discipline)
  • Offline legal / regulatory corpus
  • Connectors to internal systems (Postgres, SharePoint, iManage, FHIR, SCADA, etc.)

Capabilities that don’t work

  • External LLM providers (by definition — they require internet)
  • Cloud-based SaaS connectors (Notion, Slack cloud, Google Workspace cloud)
  • External APIs (TripAdvisor, Skyscanner, commercial market-data APIs)

Everything on that list can be substituted by internal equivalents where needed.

Certifications & compliance

Our air-gapped architecture is designed to meet:

  • Czech NBÚ (National Security Authority) requirements for classified information systems
  • German BSI (IT-Grundschutz) baseline and advanced levels
  • EU NIS2 operator of essential services requirements
  • ISO 27001 Annex A controls for segregated networks
  • IEC 62443 for industrial control systems

Customer-specific certification (e.g. for defence contracts) is a collaborative effort — we provide the architecture, documentation, and technical support; you run the certification process with your authority.

Deployment process

  1. Security review with your internal team + ours
  2. Infrastructure sizing (GPUs, CPU, storage)
  3. Package delivery via your controlled transfer process
  4. On-site or remote (through your approved channel) installation
  5. Identity integration (SSO via your internal IdP)
  6. Connector configuration
  7. Security certification / audit
  8. Go-live

Timeline: typically 4–12 weeks depending on your security review complexity.

Support model

Because we can’t reach your deployment remotely, support is different:

  • Named engineers assigned to your deployment
  • Health-check tools bundled — you run them, share outputs with us through your approved channel
  • On-site support available for incident response
  • Scheduled security briefings for your team

Related

Previous On-Premise Next Hybrid