How to Build an Air-Gapped Private Coding Setup

Keeping your code, prompts, and AI responses entirely off the internet is no longer a research-lab fantasy. An air-gapped private coding setup runs a capable model locally through Ollama, with every network path cut, so nothing you type or generate ever leaves the machine. The catch is sequencing: you have to pull and cache every model weight while you are still online, because once the network is gone there is no fetching anything.

Quick Answer

You build it in two stages. First, while connected, install Ollama and pull your coding model (for example a 14B or 24B coder) so the weights sit in the local model store. Then disconnect the machine completely and point your editor at the local Ollama endpoint. With the weights cached, inference runs with no API calls, no authentication checks, and no telemetry.

What air-gapped actually means here

Air-gapped is stronger than "private cloud" or "self-hosted with a firewall". It means the workstation has no live route to the internet at all, so even a misconfigured tool cannot phone home. For sensitive client code, regulated work, or anything under a strict non-disclosure obligation, that hard boundary is the point. The trade-off is that you take on everything the cloud used to handle: model storage, updates, and the compute to run inference.

Step 1: Pick hardware that can actually run the model

Local models are memory-bound, so plan around RAM and GPU VRAM rather than clock speed.

• A 14B coder model at 4-bit quantisation wants around 32GB of system RAM and a 16GB GPU to stay responsive.
• A 32B coder model is more comfortable with 64GB of RAM and a 24GB-class GPU.
• Smaller 7B to 8B models will run on 16GB of RAM, which is the realistic entry point if you are starting out.

If you are speccing a fresh machine for this, the AI-ready PCs at Evetech cover exactly these high-VRAM, high-memory configurations.

Step 2: Install and pull everything while online

Do all of this on the connected machine before you cut the cord.

1. Install Ollama from its official release.
2. Pull your chosen coding model, for example a current 14B or 24B coder, so the weights land in the local store under your Ollama models directory.
3. Pull a smaller fallback model too, so you have a lighter option for quick edits.
4. Install your editor and the local-agent extension you plan to use (a VS Code agent layer that talks to Ollama is the common choice), and confirm it works against the local endpoint while still online.
5. Run a test prompt end to end so you know the whole chain functions before anything is removed.

Step 3: Transfer to the air-gapped machine (if separate)

If your secure workstation is a different box that has never been online, copy the Ollama installer and the entire local models directory onto an encrypted external SSD, then move them across. On the target machine, install Ollama from the copied installer and drop the model files into the matching directory. Ollama will recognise the cached weights without needing to re-download anything.

Step 4: Cut the network and lock it down

Disable Wi-Fi and unplug the Ethernet. For belt-and-braces assurance, block outbound traffic at the firewall so even a future mistake cannot leak data. Point your editor's model setting at the local Ollama address and confirm completions still stream. Because everything is local, there are no rate limits and no per-token cost, which changes how freely you can lean on the assistant.

Living with an offline setup

The honest limits are worth stating. Local 14B to 32B models are strong for everyday completion, refactors, and multi-file edits, but they will not match the largest cloud models on the hardest reasoning. Updates are manual: when a better model lands, you repeat the online-pull step on a connected machine and transfer it across. If you want to compare your build against current popular configurations, the Evetech PC best sellers show the memory and GPU pairings that hold up under heavy local inference.

Frequently Asked Questions

Does Ollama really work with no internet?

Yes, once the model weights are pulled and cached locally, Ollama runs entirely offline. It performs no authentication or telemetry calls during inference, so a fully disconnected machine keeps working normally.

What is the smallest machine that can do this?

A 7B to 8B model will run on 16GB of RAM, which makes it the practical floor. It handles autocomplete and small edits, but for agentic multi-file work you will want 32GB of RAM and a 16GB GPU.

How do I update models on an air-gapped box?

Pull the new model on an internet-connected machine, copy the updated model files onto encrypted external storage, and move them to the air-gapped machine's Ollama models directory. There is no automatic update path by design.

Will a local model match a cloud assistant?

For routine coding it gets impressively close, especially purpose-built coder models in the 24B to 32B range. On the most demanding reasoning tasks the largest cloud models still lead, so set expectations around steady everyday help rather than peak capability.

Why not just use a VPN or private endpoint instead?

Those still send data over a network, which is the exact risk an air gap removes. If your requirement is that code never touches an external connection at all, only full disconnection satisfies it.