
Some networks are cut off on purpose
An air-gapped network has no connection to the internet. The NIST glossary defines an air gap as an interface where two systems "are not connected physically" and data moves "only manually, under human control."¹
Organizations keep their most sensitive systems this way so that private data and intellectual property cannot leave.
- The engineers inside still want AI coding assistants and agents working on internal systems.
- Every agent request needs to be checked and recorded, and whatever does the checking has to work with no internet at all.
- A tool that has to reach a vendor's server to start, update or check in cannot be approved there.
When the checkpoint needs the internet, the security team cannot sign it off. AI work stays on the shelf, or happens outside the controls.
What Argent does: built air-gap first
Argent Strategic Intelligence is the gateway between your AI agents and the AI models they use. It was built to run with no internet access: everything it needs runs inside your own network.
| Part | Where it runs |
|---|---|
| The gateway | Inside your network. Every agent that works through Argent sends its requests here to be checked and recorded. |
| Signed records | Each decision goes on a signed, tamper-evident record. By default the evidence archive is stored inside the install, on its own object-locked store. |
| Sensitive-data detector | Ships in the signed release with its language model built into the image, so it needs no outside lookup. |
| The console | Loads only from the install. Every build is scanned and fails if the console would request anything from another site; fonts and the world map are bundled. |
| The AI model | Can live inside too. Routes can point at local engines such as Ollama, LM Studio and vLLM, each turned on by an admin. |
| Argent's own AI | Not needed. Argent needs no AI model of its own to do its job; the few features that use one are optional, and the company picks the model, even a local one. |
How it works, step by step
An engineer's AI agent, inside a bank's sealed network, asks a local model to summarize an internal file.
- The request goes to Argent. The agent is pointed at the gateway, so its request arrives there first, from inside the network.
- Checked. The gateway knows which agent sent it and which person owns that agent, and checks the request against your policy.
- Passed to the local model. An allowed request goes on to the model running on your own servers, such as Ollama. The request stays on your own network.
- Recorded. The gateway writes a signed record of the request and the decision.
- A risky request is stopped. When a request carries sensitive data, such as a customer's Social Security number, sensitive data redaction catches it and the request is blocked before it reaches the model, with its own signed record.
- Same for agent frameworks. Agents built with frameworks like LangGraph work through the gateway the same way, with a short-lived agent key and no provider key.
Installed and updated by media you control
Nothing is downloaded inside the sealed network, and there is no update service. The release arrives as one signed bundle on removable media, and your own team runs the checks on your side.
- Carry it in. The bundle holds every container image, with signatures, software bills of materials and build provenance for the Nyedis-built ones, plus a checksum list, the import script, the Helm chart and the single-server bundle.
- Check the signature. The import script checks the release signature against your own copy of the Nyedis release key, or its published fingerprint, delivered separately from the bundle. It refuses to run without one.
- Check every file. The signed release must pin this bundle's checksum list, and every file must match it.
- Check every image, then load it. Each image must match the digest the signed release pins. It is copied into your own registry, and every Nyedis-built image's signature is verified there. Any mismatch prints
FATAL:and stops. - Install from your registry. One setting points the install at it:
global.imageRegistryfor the Helm chart, orARGENT_REGISTRYfor the single-server bundle. Per-install secrets are generated on the box and never printed. - Update the same way. The next release arrives as the next signed bundle and goes through the same checks; the database migrates itself on start.
What leaves the network, and who decides
NIST SP 800-53 control SC-7(5), part of the US federal moderate and high security baselines, says to deny network traffic by default and allow it by exception.² Argent is built to work behind that posture: its own services open no outside connection by default, so your perimeter can deny the rest, and anything that leaves is something your team routes or switches on.
| Connection | Default |
|---|---|
| Agent requests to an AI model | Go only where your routes point: a local model inside the network, or an outside provider you approve. |
| Model price feed | Off. Runs only if an operator turns it on, and even then new prices wait for an operator to approve them. |
| Connectors (identity directory, alerts, vault) | None at install. Each exists only when an admin adds it, and an allowlist can make connector traffic deny-by-default. |
| Console | Loads nothing from another site. The world map is drawn offline. |
| Certificates and tracing | A local certificate, no public certificate authority. Traces go to the collector inside the install. |
| Evidence archive | Stored inside the install. A cloud object store is used only if you change the setting. |
Where the company lets a connection out, the same gateway also reaches outside AI providers: AWS Bedrock, Azure AI Foundry, OpenAI, Anthropic, Google Gemini, OpenRouter and more. In all, Argent works with more than four thousand models today, and each request is checked on its way out. See model control.
What the operator sees
On the import host, a plain terminal readout. In the console, the same Argent the rest of the company uses.
- The import: one banner per check, from
== 1. release signatureto== 4. every image is the pinned digest; load it, one line per image marked as verified, and a finalOK: bundle verified and loaded intoyour registry. A tampered copy stops the run: a changed file is listed asFAILED, and a changed image or signature prints aFATAL:line naming it. - The single-server install: a host check list ending
PREFLIGHT OK, orPREFLIGHT FAILEDwith the check that failed. - The console: opens from the install with no outside request. The installed nodes are listed under Platform > Nodes.
- Each request: its signed record, allowed or blocked, with the agent, its human owner and the reason. Local models are metered too, at a cost of zero.
- The evidence: hand it to an auditor with one-click audit evidence, or replay what an agent did with session replay.
Runs where you need it, grows with demand
Argent runs as containers on Kubernetes, a standard way to run software, so the same signed release installs in AWS, Azure, Google Cloud or on your own servers.
- One Helm chart for Kubernetes 1.34 or newer, with profiles for Amazon EKS, Azure AKS, Google GKE and self-hosted clusters such as k3s and RKE2. The self-hosted profile runs the whole data tier inside the cluster: database, cache, secrets store and the object-locked evidence store.
- A single-server bundle for one Linux host, with a host check and per-install secrets.
- Can grow on demand. On Kubernetes with autoscaling turned on (the production profile turns it on; the cluster needs KEDA), when more agents send requests, Argent starts extra copies of its gateway side by side and shares the work between them. When things quiet down, the extra copies go away on their own. The production profile keeps at least three running.
- Measured install times. In a fresh AWS account, eleven runs from start to every service ready took 17.7 to 35.1 minutes, a median of 28.9. On one cloud server, the single-server bundle was serving 1.6 to 3.5 minutes after unpacking, pulling its images from a private registry. An air-gapped install adds the time it takes to carry in and load the bundle.
Compare air-gapped on premises, private cloud and public cloud VPC on the deployment options page.
Where it fits, and how a team starts
Air-gapped install is the base of sovereign AI: your AI governance runs on your network, under your rules. See the sovereign AI goal.
- Use cases: AI in regulated and air-gapped environments, AI coding agents, enterprise copilots, MCP and agent-to-agent tool access, customer-facing AI assistants.
- Goals: see all AI in your company, stop data loss, audit and accountability, compliance evidence, model control and routing and all goals.
- Industries: banking, public sector, military and defense, healthcare, insurance, utilities and all industries.
- Carry in and check. Bring the signed bundle across, run the import, and load your own registry.
- Install and route. Point the chart or the single-server bundle at your registry, then point your agents at the gateway and your routes at local models.
- Watch, then enforce. Start in watch mode, which records what Argent would have done while nothing is stopped, then turn on enforcement. Take the walkthrough or browse all capabilities.