AI sovereignty: why owning your stack matters now

Every AI product asks the same thing of you, and it is rarely stated plainly. Send us the conversation. Send us the documents you are reasoning over. Send us the half-formed idea you have not shown anyone yet. In exchange, you get a good model and a fast interface.

That is a real trade, and sometimes it is the right one. What is missing is the option to decline the parts you did not agree to.

Convenience is not the problem

It is tempting to argue this as a scare. Your data is being harvested, the model is training on your work, one day the terms will change. Some of that is true some of the time, and none of it is why sovereignty matters.

The real reason is duller and harder to dismiss: software you do not control can change underneath you. The model you built a workflow around gets deprecated. The price triples. The context window shrinks to protect margins. A feature you depend on moves behind a higher tier. The company is acquired and the roadmap becomes someone else's.

None of that requires bad intent. It only requires that the decisions live somewhere other than with you.

What sovereignty actually means

Not isolation. Not refusing cloud models. Not running everything on a machine under your desk out of principle.

Sovereignty is narrower and more useful than that:

• The workspace runs where you decide. Your laptop, your server, a VPS you rent, a machine in a hospital basement with no route to the internet. The software should not care.
• The pieces are replaceable. If a provider disappoints you, you change the route, not the tool. The conversation, the documents, and the history stay put.
• The record is yours. Your chats, your files, your agent runs live in storage you can back up, inspect, encrypt, and delete without asking anyone.
• The claims are checkable. "We do not track you" is a marketing line until the source is open and someone can go look.

That last one is the load-bearing part. Every privacy promise made by closed software is, in the end, a promise about intentions. Open source turns it into a statement about behaviour.

Local first, cloud by choice

The version of this argument that fails is the purist one: run everything locally, refuse every API, accept a worse model as the price of principle.

Most work does not survive that trade, and it is not necessary. A local model handles far more than people expect — drafting, summarising, extraction, classification, most day-to-day reasoning over your own documents. When you genuinely need a frontier model, use one. Deliberately, for that task, knowing what leaves the machine.

The distinction is not local versus cloud. It is default versus choice. A tool that is local by default and remote by choice leaves the decision with you. A tool that is remote by default has already made it.

What this looks like in practice

This is the thinking behind Libre WebUI, so take the following as an argument with a product attached rather than a neutral survey.

It runs on your hardware. One command and you have a workspace at localhost:8080 — no account, no telemetry, nothing phoning home. Ollama gives you local models; provider connections let you route specific work to OpenAI, Anthropic, Google, Mistral, Groq or any OpenAI-compatible endpoint, per conversation, when you decide it is worth it.

Documents you attach are chunked and embedded on your machine. Notes and knowledge collections are encrypted at rest in your own database. Work gives an agent a sandboxed container with files, a terminal and a preview — and the container is bounded, non-root, read-only at the root filesystem, with the workspace volume as the only writable mount. If you switch it off, nothing was ever running elsewhere.

The commitments are written down rather than implied. Apache 2.0 in perpetuity, with no relicensing to source-available or proprietary terms. Zero telemetry. No VC equity, because accepting it would eventually make someone else's growth the reason a decision gets made. The stewardship is named — Kroonen AI, Inc. — so there is somewhere for responsibility to sit.

The part people skip

Sovereignty has a cost and it is dishonest to pretend otherwise.

You will maintain something. You will pick models instead of having one picked for you. You will occasionally read a log. A hosted product hides all of that, and for many people that is a fair exchange — it is a legitimate choice, not a failure of nerve.

But the cost is smaller than it was even a year ago, and it is falling. Local models that were research toys are now genuinely useful on a laptop. Self-hosting that meant a weekend now means one command. The gap between "I would like to own this" and "I can own this" has closed enough that it is worth asking which side of it you are on.

Why it matters more now

Because the defaults are hardening.

AI is moving from a thing you visit to a thing embedded in how you work. It reads your documents, drafts your writing, runs commands in your projects, remembers your context across months. Every one of those is a place where a dependency becomes structural instead of casual.

Dependencies are easy to accept while they are still small. They are difficult to unwind once your work is shaped around them. The moment to decide where your stack lives is before that shape sets — not after a pricing page changes.

You do not have to run everything yourself. You should be able to.

---

Try it: npx libre-webui and open localhost:8080. Quick start · Source on GitHub · Why Libre