Skip to content
Guides & References

Reference design: an air-gapped AI NOC with rConfig, a self-hosted LLM and local agents

A fully on-site AI design: rConfig, monitoring, a self-hosted model (vLLM or Ollama) and self-hosted agents, with no traffic to a cloud AI provider.

Iris: AI CenterReference designUpdated
On this page

Reference design: an air-gapped AI NOC with rConfig, a self-hosted LLM and local agents

Some networks can’t talk to a cloud AI provider at all: OT and industrial sites, defence, critical infrastructure, classified enclaves. They can still use AI on their configurations. In this design every component runs on site: rConfig, monitoring, the model, and the agents that use it.

What it looks like

flowchart LR
    subgraph site["Site network (no internet)"]
        RC["rConfig + Iris"] --> M["Model server<br/>vLLM or Ollama"]
        N["n8n<br/>AI Agent"] --> M
        N -- "MCP" --> RC
        N -- "MCP or API" --> MON["Monitoring<br/>Zabbix / LibreNMS"]
        RC -- "webhooks" --> N
        N --> CH["On-site chat<br/>or ticketing"]
    end
Component Role
rConfig with Iris Config backups, compliance, and Iris’s features, all on the local model
Model server One self-hosted model, shared by Iris and the agents. vLLM if several workflows run at once
n8n Self-hosted agent workflows: webhook-driven narration, digests, a chat box. Its AI Agent node uses a chat model node pointed at your local server
Monitoring Health and alerts, read by the agent

What changes compared with a connected site

  • No cloud agents. Claude Code, Cursor and Copilot use their vendors’ models. Use agents you can point at your own model server: n8n’s AI Agent, or an agent framework configured for an OpenAI-compatible endpoint.
  • One model serves everything. Size it for the busiest moment: Iris users and every workflow at once. See Sizing a local model.
  • Answers are bounded by the model. Keep agent tasks narrow and well defined (explain this diff, summarise this group) rather than open-ended investigations.

Workflows that fit

Frequently asked questions

Can I use AI agents with rConfig on a network with no internet access?

Yes, if the agent and its model both run on site. A self-hosted model server (vLLM or Ollama) serves Iris inside rConfig and a self-hosted agent such as n8n's AI Agent, which connects to rConfig over MCP.

Can Claude Code, Cursor or Copilot be used in this design?

Not as usually deployed. Their models run in the vendor's cloud. This design uses agents that can be pointed at a model on your own network.

What does the site need from outside?

Nothing at run time. Models, container images and updates are brought in through your normal controlled transfer process.

Type to search.