Air-gapped AI for government and critical systems
Air-gapped AI keeps the model, platform, identity, evidence and operational dependencies inside a disconnected environment. Acceptance requires testing installation, licensing, updates and recovery without external services. A container bundle alone does not demonstrate a complete offline deployment.
Evaluating this for your team? Discuss your use case, deployment boundary and success criteria in a technical briefing.
Book a technical briefingDefine the disconnected boundary
List every dependency before choosing a deployment. Include the model server, licence checks, identity provider, image registry, package sources, DNS, time, certificates, telemetry and operator access.
Agree the release and acceptance procedure
Customer-hosted deployments use packaged deployment configuration. An offline deployment requires an accepted bundle, local model and disconnected installation tests.
Connected self-hosting and disconnected operation require different acceptance evidence. Agree the offline licence, import process, trust key and complete local dependencies before committing to an isolated rollout.
Keep inference and identity inside
Bring your own model — OpenAI-compatible endpoints, Ollama, Anthropic, or Cloudflare AI Gateway / Workers AI — with keys envelope-encrypted at rest.
Customer-hosted installations support an identity provider through OIDC with PKCE. Hosted workspace SSO must be confirmed for the agreed deployment.
For a disconnected design, select a local model and local identity service; a hosted endpoint remains an external dependency.
Test operations with the network disconnected
- Install, enrol a supported host and complete the agreed workflow.
- Verify that no external connection is required for licensing or sign-in.
- Verify package authenticity and test a rejected tampered package.
- Restore a backup and test an approved update and rollback.
Questions buyers ask
Is air-gapped the same as on-premises?
No. A customer-hosted deployment may still use connected licensing, registries or hosted inference. A disconnected design must remove these dependencies.
How do we accept an offline deployment?
Test disconnected installation, local inference and identity, offline licensing, controlled updates and backup restoration against a named release.
Customer-hosted deployments use packaged deployment configuration. An offline deployment requires an accepted bundle, local model and disconnected installation tests.
Can the model run locally?
Use a local model endpoint with weights and serving dependencies already available inside the approved environment.
Bring your own model — OpenAI-compatible endpoints, Ollama, Anthropic, or Cloudflare AI Gateway / Workers AI — with keys envelope-encrypted at rest.
How do updates enter the environment?
Use the customer-approved media-transfer process and verify the accepted package before installation. Test authenticity, tamper detection and rollback as part of acceptance.
Where to go next
All guidesBring one question of your own.