Sovereign cognitive AI that runs on your own hardware.
Cognitive OS is a full AI system, not a wrapper around a chatbot. The language model is just one component — goals, memory, decisions, and self-repair live outside it. It runs on your own infrastructure, keeps your data where it lives, and tells you when it doesn't know instead of making things up.
Hub71 startup · Abu Dhabi, UAE · Built on Active Inference
See the whole system in four minutes.
Cloud AI asks you to give up three things.
Your data flows to servers you don't control.
Your uptime depends on someone else's cloud.
Your privacy becomes a setting, not a guarantee.
The model is a component. Not the mind.
Most agents are a language model stuck in a loop. Cognitive OS flips it: goals, beliefs, action-selection, memory, and self-repair are explicit and inspectable — and they live outside the model. The LLM is a tool the system uses, not the thing it is. That's what makes the system auditable and recoverable.
It's built on Active Inference (Friston): the system keeps a model of its world, acts on what it knows, and explores where it's uncertain.
An LLM answers. Archotec keeps working.
An LLM is a hammer. Archotec is the builder that picks it up.
LLMs think once. Archotec thinks over time — it works while you sleep.
Need a quick answer, or a snippet? Use an LLM. Need something researched and understood over weeks? That's Archotec.
A system built around control and honesty.
Runs on your infrastructure
Everything runs on hardware you control. Your data stays where you run it.
It doesn't make things up
When it doesn't know, it says so or grounds to a source. It won't bluff a confident wrong answer.
A system, not a prompt
Memory, goals, and an honesty layer wrap the model — not a clever prompt.
Builds from a description
Describe a system in plain language; it designs it, tests in a sandbox, assembles what works — and tells you what it couldn't.
Auditable by design
Every decision traces to explicit logic. When a part misbehaves, it can roll back and explain why.
Persistent memory
It remembers across sessions — a real world-model, not a context window.
On a small internal honesty test, the same raw model answered honestly about 30% of the time. Inside Cognitive OS, about 99% — on the kinds of questions models tend to fabricate. It's an early, self-graded benchmark, not an independent eval.
In a recent run of 5 real build goals: 4 completed with real, working output, 1 failed honestly with no fabricated result, 0 hallucinations.
On harder, adversarial test sets the honesty rate is lower — around 90%+ — with some residual misses on technical edge cases. We report both numbers openly rather than only the flattering one.
In our testing: 0 jailbreak or manipulation prompts were accepted as tasks. The system held its safety boundaries on every attempt. These are our own tests, not independent certification.
Give it a goal. Watch it build.
And if it can't finish something, it tells you — instead of pretending it did.
Built for places that can't put their AI in a foreign cloud.
It's an early alpha. On purpose.
Cognitive OS is in early access — powerful, and still rough in places. We're opening it to a small group by access key, and we want your honest feedback. No hype, no perfect demo. Just the real thing, early.
Request early access
Tell us a bit about you and we'll send an access key.
Common questions
What exactly is Cognitive OS?
It's a full AI system where the language model is one bounded component among several. Goals, memory, action-selection, safety checks, and self-repair are explicit modules that live outside the LLM — making the system auditable, recoverable, and honest about what it doesn't know.
Does it run on my own hardware, or on your servers?
The product is designed to run entirely on your own infrastructure — nothing leaves your network. The current alpha is different: it's hosted-by-key, meaning your inputs run on our instance during the trial. We tell you this upfront. On-premise deployment is the goal and the product; the hosted alpha is a way to let you try it before standing up your own stack.
Is my data private in the alpha?
In the alpha, your inputs are processed on our hosted instance, so full on-device privacy doesn't apply yet. We don't store or share your data, but we're being precise: 'sovereign' and 'on-premise' are properties of the product, not of the hosted trial. If that distinction matters for your use case, we can discuss on-premise deployment directly.
How is this different from a fine-tuned LLM or a RAG pipeline?
A fine-tuned LLM is still a black box that can hallucinate. A RAG pipeline adds retrieval but the decision logic is still inside a prompt. Cognitive OS makes goals, beliefs, and decisions explicit and inspectable outside the model. The LLM generates and retrieves — it doesn't decide.
What does 'built on Active Inference' mean in practice?
Active Inference (Friston) gives the system a principled way to handle uncertainty: it maintains a model of its world, acts on what it believes, and explores where it's uncertain. In practice it means the system can say 'I don't know' in a structured way, not just as a fallback phrase.
What's the current state — is it production-ready?
No. It's an early alpha — functional in constrained settings, honest about failures, and still rough in places. We're opening it to a small group who want to build with it and give real feedback. We're not claiming production-readiness; we're claiming it does what it says and tells you when it doesn't.
How do I get access?
Fill in the request form on this page. We read every request and send access keys to people whose use cases fit what we're building toward. Early access is limited.
Have more questions? Contact our team