No internet. No problem.

Most AI solutions you meet today are a thin app in front of a remote service. You type something, it is sent away, a model answers, and the answer comes back. It works flawlessly — until the connection drops, the subscription changes price, or somebody asks where the data actually sits.

We chose the other road. The agents run on a machine standing in the building. The model is a file on the disk. Documents, tools, logs and decisions all stay in the same place. Nothing is called out over the network, because there is nothing to call: everything the agent needs, it already has locally.

What a cloudless agent is

A language model of a size that fits on an ordinary graphics card or a small workstation — typically between eight and seventy billion parameters — combined with what turns a model into an agent: a set of tools, a workflow, a log, and a number of points in the flow where a human has to approve the next step.

The architecture is the same as in the cloud. Only the location differs. A typical setup at a customer is one machine under a desk, drawing around 200–400 watts under load, and keeping at work when the phone has no signal, the factory network is down, or somebody fiddles with the router over the weekend.

Why we chose it this way

• The data stays in the building. Production data, personal data and contracts do not leave the address. That makes data processing agreements shorter and the conversation with a security officer considerably easier.
• No dependency on a third party. No vendor can raise the price, change the terms, retire the model you built on, or shut off access overnight.
• Predictability. The model does not change behind your back. The version that was approved is the version that runs — six months from now as well.
• Fixed cost. You buy hardware once instead of paying per token forever. At steady daily use that is often the cheaper route.
• It works without a network. On the factory floor, in the summer house, in a field, on a ship, in a basement room. Places where a cloud service is a poor idea in practice.

What we do today

We use what we sell. Our own operation runs on local models: agents that read and sort documents, write minutes, track tasks, monitor workflows and ask permission when an action is sensitive.

• Field extraction from invoices, delivery notes, quotes and PDFs — into spreadsheets or a case system.
• Meeting notes and minutes, including speech-to-text in Danish.
• Email triage: what to answer, what to file, what is urgent.
• Assembling reports and status from data that is collected by hand today.
• Monitoring a workflow and alerting when something deviates from the normal.
• Several agents sharing the work, with a visible panel of what is happening right now — and a human as the decision point.

What we can build for you

We do not start with a technology, but with a task. The process usually looks like this:

1. Mapping. One or two days at your site or by video. We find the tasks that steal the most time and that lend themselves to automation.
2. Bounded trial. Two to four weeks with one task, your real data and a success measure agreed up front. Not a demonstration — a measurement.
3. Decision basis. An honest split: what can be solved locally, what still needs a cloud service, and what should remain manual work.
4. Integration. We connect the agent to what you already use — ERP, email, calendar, case system, spreadsheets or a shared drive — so data is not entered twice.
5. Security and operation. Access rights, approval points, logging and a machine-checked gate before going live. Then monitoring, updates and a fixed contact.
6. Training. The people who will work alongside the agent learn what it can do, what it cannot do, and how to stop it.

What cloudless is not

It is not magic, and it is not free. A local model is not the largest model. Very broad world knowledge and the hardest language tasks are still handled better by the large cloud models, and there are tasks where sending data away is the right solution. Local means hardware with enough memory — and someone to maintain it.

We say so when that is the case. An agent running locally is not automatically secure — it is only secure if somebody has set limits on what it may touch, and verified that the limits hold. That work is not something you buy your way out of with a model.

The point

Cloudless does not mean old-fashioned. It means the AI can be in operation without data leaving the building, and without anyone being able to shut it off with a price increase or a failed API. The question is no longer whether the AI may see your data. It is whether it has to take it with it.

Would you rather see it run than read about it? We will gladly put a machine at your site and solve a real task on it.