AIgapped Request early access

Modern AI inside networks that never touch the internet.

AIgapped packages open-weight LLMs, the runtime and every dependency into one signed bundle. It crosses your air gap on a drive or data diode and installs with one command. Updates arrive the same way.

Request early access See how it works

Connected build side

  • Open-weight models
  • Container images
  • CVE scans and SBOMs
  • Evals and signatures

Your network

bundle aigapped-2026.11 signed
Contents of an example bundle
model-weightssha256:9f2c…e41averified
inference-serversha256:41bd…07c3verified
chat-uisha256:c88e…9a10verified
document-searchsha256:7a03…d2f8verified
kubernetes-runtimesha256:e519…3b6everified
sbom.spdx.jsonsha256:0d7f…b4c2verified

0 outbound connections

An AIgapped bundle is built and signed on the connected side, carried across the air gap, and verified inside your network before anything runs.

Everything assumes the internet is there

Running an open model is easy on a laptop with Wi-Fi. Inside a disconnected network, the commands every tutorial depends on simply fail. Each one becomes a manual export, a transfer request and a security review.

AIgapped does that work ahead of time, on our side of the gap, and hands your team one bundle your security reviewers can check line by line.

How it works

  1. Build

    On the connected side we pull the models and every container image the stack needs, scan them for vulnerabilities, generate an SBOM, run evals and sign the result.

  2. Cross

    The bundle moves through your existing transfer process: encrypted drive, one-way data diode or whatever your policy allows. Nothing needs to reach back out.

  3. Install

    One command verifies every signature, then brings up Kubernetes, inference, the chat interface and document search on your hardware, GPU or CPU.

  4. Stay current

    New open models and security patches ship as signed update bundles through the same path, so you aren't stuck on whatever you installed first.

What runs inside your walls

Models
Open-weight LLMs such as Llama, Qwen, DeepSeek, Mistral and gpt-oss, chosen and evaluated for your hardware.
API
An OpenAI-compatible endpoint, so internal tools and existing code work without changes.
Chat
A web interface your staff can use on day one.
Document search
Ask questions across internal files. The documents and the index never leave the network.
Access control
Single sign-on, roles and an audit log of every request.
Runtime
Lightweight Kubernetes on a single box or a multi-node cluster.
Evidence
Signatures, SBOMs and scan reports for each bundle, ready for your security review.

Who it's for

Regulated organizations

Defense suppliers, banks, hospitals, utilities and government teams whose most valuable data can't be sent to an AI API.

Software vendors

AI products whose customers love the demo, then ask for it to run entirely inside their building. Ship your product with the models and runtime it needs.

Hardware and integration partners

Server builders and integrators who want a ready, supportable AI stack on the machines they deliver.

Why I'm building this

When DeepSeek-R1 came out, I deployed it on a three-node Kubernetes cluster at home, with no cloud, and wrote up how. The manifests are on GitHub.

The model was the easy part. Storage, scheduling, a 4.7 GB download and secure access were the real work, and every step still assumed internet access. Take that away, as banks, hospitals and defense networks must, and nearly all of it breaks. AIgapped is the version that works when the cable is unplugged.

Akash Awase, founder

Request early access

AIgapped is in early access. We're working closely with a small group of design partners on their first disconnected deployments.

Tell us about your environment and we'll reply within two business days. We use your details only to contact you about AIgapped.

Which describes you best?