Your data stays on your machines

Sovereign AI: agents that work on your premises

Many businesses give up on AI agents for one reason only: they do not want to upload their customer files, quotes and correspondence to a platform whose walls they cannot see. nullbot starts from the other side. The software installs on the company's own computers, on Mac or Windows, and what your agents handle stays stored there. You then choose, agent by agent, which AI model does the thinking — down to a model that runs on the machine itself. And because the team lives on your premises, it can look after your files for months on end without their contents changing hands.

1. What caution costs

A business that is careful with its data pays for that care in another way. The practice that refuses to expose its clients' files keeps handling everything by hand. The workshop that protects its drawings delegates nothing that touches them. The shop that hesitates to share its customer list lets its follow-ups slip. The bill never says “security”: it takes the form of hours, postponed tasks and a position that will never be opened.

The usual answers move the problem without solving it. An online tool asks you to trust hosting you do not control. A custom build, installed on your premises, costs a whole project and needs someone to maintain it. What was missing was a team that installs like ordinary software and works where your files already are.

2. Where your data lives, and what leaves

Sovereignty is not something you declare; it is something you check. Here, without rounding, is what stays and what goes.

What stays on the machine: customer records, conversations with your agents, the documents they produce, the history of their tasks and the log of what they have done. Everything is stored in a database installed with the software, on the computer where it runs.
Your credentials: the logins for your tools and your AI model keys are kept on that same workstation. You can withdraw or replace them at any time with the services concerned.
What our servers see: the management of your account and your licence. They are not a store for the content of your agents' work.
What leaves because you connected it: an agent that reads your mailbox or posts on a network necessarily exchanges with that service. Every connection is declared, can be removed in one step, and nothing is added behind your back.
What goes to the AI model: with an online model — nullbot AI, your Claude or ChatGPT subscription, an API key — the text handed to the agent goes to the provider running that model. That is the passage a local model removes.

3. The model that runs on your computer

In the desktop app, nullbot can install and run an open AI model by itself. The model is downloaded to disk once, then answers from the computer, with no subscription and no usage bill. The catalogue brings together several well-known families — Qwen, Gemma, Ministral, Granite — in different sizes, from the lightest, comfortable with 4 GB of memory, to the largest, which need about 24.

The choice is made agent by agent. The one sorting sensitive documents can work with the local model, while another, in charge of a public text, keeps a more powerful online model. A built-in test shows you which model actually answered, and only the administrator of the installation can add or remove one.

4. Week 1, month 3, year 2

An installation that keeps data in-house is only worth something if it really works. The difference shows over time.

Week 1: the software is installed, the first agents receive their assignments, and you decide which tools they may open. The local model is tried on a concrete task to judge whether it is enough.
Month 3: the agents have settled into your real files — follow-ups, sorting requests, preparing documents. Every action stays in the log, and the budgets set per agent cap what the online models cost.
Year 2: the team's history — customers, exchanges, decisions — belongs to the company and sits on its machines. Changing model or provider loses nothing: the memory of the work was never with them.

5. What to know before choosing

A local model has its limits, and it is better to know them in advance. It needs a sufficiently powerful computer, and it remains less at ease than a large online model on long or delicate tasks. Many businesses therefore combine the two, keeping the local model for content they do not want to see leave.

The free choice of model, local included, comes with the Multi Pro and Finance plans; the Solo plan works with the included nullbot AI. The local model exists only in the desktop app.

Keeping your data in-house also means keeping the responsibility. The agents work while the computer is on, and protecting the workstation — locked session, encrypted disk, backups — remains your job, as for the rest of your files.

6. Frequently asked questions

Is nullbot an online platform?

No. It is software that installs on the company's computers. Our servers manage your account and your licence, not the content of your agents' work.

Can you use AI agents without sending your data to an AI provider?

Yes, with a local model, which answers from the computer itself. With an online model, the text handed to the agent goes to that model's provider: it is your decision, agent by agent.

Do you need a special computer for a local model?

The smallest models in the catalogue make do with 4 GB of memory; the largest need about 24. A good starting point is to try a mid-sized model on a real task before using it more widely.

What happens to my data if I stop using nullbot?

They stayed on your machines throughout your subscription. Stopping nullbot brings nothing back from anywhere, since it was never stored with us.