What are air-gapped and on-prem AI agents?
On-prem AI agents run in your own data centre; air-gapped AI agents run in a site with no external network at all. In both, the agent's model, its tools and its governance (policy, approvals, the kill switch and the audit trail) run inside the boundary, on infrastructure you control, with keys only you hold. Nothing calls home, and the system keeps working with no internet connection. This is the strictest form of sovereign AI for agents.
What does sovereignty mean for on-prem AI agents?
For a chatbot, sovereignty is mostly about where the model and the data live. Agents add actions: they hold credentials, call tools and change systems. So the governance layer needs the same sovereignty as the model. Five questions cover it:
- Where does it run? The agent, its model, its tools and its governance all run on infrastructure in a known place. A self-hosted model behind a hosted policy service is only partly sovereign.
- Whose keys? The keys that sign agent identities and audit checkpoints should be yours. Whoever holds them can mint identities or re-sign records.
- Which law applies? Data held by a provider can be subject to that provider's home law. For example, the US CLOUD Act lets US authorities seek data held by US-based providers even when it is stored abroad (US Department of Justice).
- Does it phone home? Telemetry, update checks and licence calls all send data out and create a dependency. A sovereign system makes none.
- How is it licensed? A licence that must be checked online fails in a disconnected site. It should verify offline, and an expired licence should never stop agents mid-task.
Evidence matters too. If your audit trail is shipped to a vendor's region, part of your oversight now sits outside your jurisdiction.
Which deployment topologies fit on-prem and air-gapped AI agents?
| Topology | Where governance runs | Who holds the keys | Internet needed | Fits |
|---|---|---|---|---|
| Hosted governance service | The vendor's cloud | Usually the vendor | Yes | Fast pilots with no residency demands |
| Self-hosted in a public cloud region | Your account, in a region you pick | You, if you bring your own | For model and tool calls you allow | Most enterprises |
| Sovereign or national cloud | An in-country provider | You | Usually | Public sector and regulated firms |
| On-premises | Your data centre | You | Optional | Banks, critical infrastructure |
| Air-gapped | An isolated site with no external network | You | No | Defence, classified and safety-critical sites |
| Confidential computing | Any of the above, inside hardware-encrypted VMs | You, with keys released only after attestation | Depends on the base | When the infrastructure operator is not trusted |
These combine. An air-gapped site can also run confidential VMs, and an on-premises installation can be highly available. For a wider comparison, see self-hosted vs hyperscaler.
How do you govern air-gapped AI agents?
In an air-gapped site, nothing reaches the internet. That rules out any governance that needs a cloud service, and it creates three practical problems.
- Getting software in safely. Images, charts and updates arrive on removable media. Each delivery needs checksums and a signature you verify with a key you obtained separately, plus an SBOM so you know what is inside.
- Licensing without a server. The licence must be a signed file verified locally.
- Models and tools inside the boundary. The agent's model and every tool it calls must run on site. Egress rules should make that true by construction, not by convention.
The governance itself is unchanged: policy on every tool call, human approval for risky actions, a kill switch that works locally, and a tamper-evident audit trail stored on site.
What should you ask a vendor about sovereign agent governance?
- Does any component make an outbound call by default? Telemetry, updates, licence checks?
- Is there a vendor control plane, and what stops working without it?
- Who holds the identity-signing and audit-signing keys? Can you use your own?
- Can you restrict which registries, model endpoints and tool hosts agents may reach, by jurisdiction?
- Is there a documented air-gapped install, with signed bundles you can verify offline?
- What happens to running agents when the licence expires?
Where MAQPNA fits
MAQPNA is a self-hosted runtime for AI agents by Kosa8 Labs. In its docs, "sovereign" means you can run agents in your own jurisdiction, on your own infrastructure, with your own keys. The same controls run in every edition; the Sovereign edition adds the hardware-backed parts (editions).
- Zero phone-home. No telemetry, update checks or licence calls. The licence is a signed file verified offline, and no licence state ever blocks or degrades agent traffic (licensing).
- A sovereignty policy. A cluster-wide
SovereigntyPolicydeclares allowed jurisdictions, registries, egress hosts and CIDRs. The operator rejects agents outside it, and the gateway dials every upstream through a residency-checking dialer.maqpna sovereignty checktests manifests offline. - Customer-held keys. The identity broker signs with your Ed25519 key, and you hold the audit checkpoint key. File-based keys work today; the docs list HSM (PKCS#11) and KMS signing backends as planned.
- Air-gapped install.
maqpna airgap bundlebuilds an offline bundle of images, chart, SBOMs and signed checksums.maqpna airgap verify --require-signature --require-keychecks it against a key you obtained separately, andmaqpna airgap pushmirrors the images into your registry (air-gapped install). - In-country evidence. Ledger segments can ship to your own S3-compatible write-once store, checked against the residency policy.
- Confidential tier. tier-2 sandboxes run in AMD SEV-SNP or Intel TDX confidential VMs. The session token is released only after the VM passes attestation against reference values you approve. The docs note that runs on real SEV-SNP and TDX hardware are not yet part of CI.
The documented topologies are local dev, single cluster, HA, multi-tenant, air-gapped and confidential. The sovereign-eu profile is HA, air-gapped and confidential at once (deployment topologies). Read more on the sovereignty page. MAQPNA holds no certifications, and installing it does not make a system compliant. See also the EU AI Act and AI agents, prompt injection in agent tool calls and stopping an agent deleting a production database.