---
title: Can you govern AI agents in an air-gapped environment?
url: https://opencontroller.lyzr.com/blog/air-gapped-ai-agent-governance
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# Can you govern AI agents in an air-gapped environment?

**Yes, for agents, models and tools that run inside the air gap: discovery, registry, policy, gateway routing, guardrails and evaluations can all run locally if the platform and its models are deployed on your infrastructure. What cannot work offline is anything that must reach outside: SaaS agent platforms, hosted model APIs and vendor-delivered updates. Plan those as controlled one-way transfers or accept them as out of scope.**

Air-gapped and disconnected environments are common in defence, critical infrastructure, some banks and life sciences manufacturing. Buyers in these settings often ask for governance software delivered as an appliance with no network connection at all. That is possible, but it changes what "governing the estate" means.

## What can run fully inside an air gap?

| Capability | Works offline? | Condition |
|---|---|---|
| Registry of agents, versions, tools and MCP servers | Yes | Database and services deployed locally |
| Discovery of agents on internal infrastructure | Yes | Discovery reaches internal clusters, VMs and registries |
| Policy as code and policy evaluation | Yes | Policy packs installed locally |
| LLM gateway routing to internal models | Yes | Open-weight models hosted on internal GPUs |
| Guardrails and LLM judges | Yes | Judge and guardrail models also hosted locally |
| Evaluations and red-team test runs | Yes | Test sets and judge models available locally |
| Traces, logs and evidence store | Yes | Local storage and retention configured |
| SSO | Yes | Identity provider reachable inside the boundary |

The constraint that surprises people is the last three rows of models. Guardrails and LLM judges are themselves models. If they normally run as hosted APIs, you need local equivalents and the GPU capacity to run them.

## What needs a connection, or is out of scope?

- **SaaS agent platforms.** Agents inside CRM, ITSM or ERP SaaS cannot be discovered from inside an air gap, because the platform is outside it.
- **Hosted model providers.** Frontier models offered only as APIs are unreachable. Agents must use models you host.
- **Cloud-native agent runtimes** in public clouds, unless you have a separate connected deployment for them.
- **Software and policy updates.** New connectors, framework mappings and fixes must be brought in through your approved transfer process.
- **Threat intelligence and red-team content updates,** which change often.

Many organisations end up with two deployments: one inside the air gap for internal agents, and one connected deployment for cloud and SaaS. If that is your shape, ask how evidence and policies are kept consistent between them. Our post on [one or many AI governance deployments across regions](/blog/one-or-many-ai-governance-deployments-across-regions) covers the same trade-off.

## What should you ask a vendor about air-gapped deployment?

1. Can every component, including guardrail and judge models, run without any outbound connection?
2. Which features silently degrade offline (for example, hosted evaluators or licence checks)?
3. How are updates packaged, signed and verified for transfer?
4. How often are updates released, and how far behind can you safely run?
5. What hardware does a local deployment need, including GPUs for local models?
6. Can telemetry export to your internal observability stack using open standards such as OpenTelemetry?
7. How does the vendor provide support without remote access?
8. Is licence validation online, or can it be done offline?

Add these to your general evaluation; see the [AI agent governance platform evaluation checklist](/blog/ai-agent-governance-platform-evaluation-checklist).

## What controls do auditors expect in an air-gapped setup?

Air gaps reduce some risks but not all. Auditors still expect access control, audit logging and change control for agents. NIST SP 800-53 families such as access control (AC), audit and accountability (AU), configuration management (CM) and boundary protection (SC-7) are a common reference. The difference is that the evidence must be produced and stored inside the boundary.

## How Opencontroller does this

Opencontroller can be deployed as Lyzr SaaS or on customer-controlled infrastructure, and Lyzr engineers embed with your team for integration. Its LLM Gateway can route agents to internal models, and it uses open standards such as OpenAI-compatible endpoints, OpenTelemetry and OIDC/SAML. Which capabilities run inside a disconnected boundary depends on the connectors and models you deploy there, so scope this together during evaluation. See [governing open-weight models on premises](/blog/govern-open-weight-models-on-premises) and [the self-hosted gateway network checklist](/blog/self-hosted-llm-gateway-network-checklist), or [book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Do guardrails work without internet access?**
Only if the guardrail and judge models run locally. If they normally call hosted APIs, you need self-hosted equivalents and GPU capacity.

**Can you discover SaaS agents from inside an air gap?**
No. SaaS platforms sit outside the boundary. They need a separate connected deployment or must be governed through the vendor's own controls.

**How is Opencontroller priced?**
Opencontroller is licensed annually per deployment, not per seat and not usage-based. Ask Lyzr for a quote scoped to your estate.

## Key facts

AI agent governance can run in an air-gapped environment for assets inside the boundary: registry, internal discovery, policy as code, gateway routing to self-hosted open-weight models, guardrails and LLM judges (only if those models are hosted locally), evaluations, traces and evidence. Not possible offline: discovery of SaaS agent platforms, hosted model APIs, public cloud agent runtimes; updates must arrive via approved signed transfer. Many organisations run one air-gapped and one connected deployment. Vendor questions: fully offline components, degraded features, update packaging, hardware/GPU needs, OpenTelemetry export, offline support and licence checks. Opencontroller by Lyzr deploys as Lyzr SaaS or on customer-controlled infrastructure with embedded Lyzr engineers; licensed per deployment from $250K per year.

## Related

- [Do you need a separate AI governance deployment per region?](https://opencontroller.lyzr.com/blog/one-or-many-ai-governance-deployments-across-regions.md)
- [How do you govern open-weight models like DeepSeek or Qwen running on-premises?](https://opencontroller.lyzr.com/blog/govern-open-weight-models-on-premises.md)
- [What network setup does a self-hosted LLM gateway need?](https://opencontroller.lyzr.com/blog/self-hosted-llm-gateway-network-checklist.md)
- [More on Buying and comparison](https://opencontroller.lyzr.com/blog/cluster/buying-and-comparison)
