Where to find an AI assistant that runs offline on your own UK servers
You can get one from Mickai, a UK company (Mickai LTD) whose AI operating system installs on servers you own and runs its AI assistant with no internet connection at all. Your staff use the assistant inside your own building, nothing is sent to a cloud service, and every consequential action is signed into a tamper-evident audit record you can check offline. If you would rather not buy a complete system, the other routes are to assemble open-source AI software and a chat interface yourself, or to take a private edition from a larger AI vendor, so the right choice depends on how much you want to build and how strictly offline you need to be.
What are your options for an offline AI assistant?
There are three realistic routes, and each suits a different organisation. The first is to build it yourself. Open-source model software and chat interfaces are capable and well supported by their communities, but your team assembles and maintains everything around them: user access, logging, updates and support. The second is a private or on-premise edition from an established AI vendor. These are strong products, so check whether 'private' means your own building with the network unplugged, or a dedicated space in someone else's data centre. The third is a complete system designed to run offline from the start. Mickai sits in this group: an operating system that installs on hardware you own and brings the assistant, access control and a signed audit record together. The right route depends on your in-house engineering capacity and how strictly offline you need to be.
What should you look for in an offline AI assistant?
Start with tests you can run yourself rather than claims on a website. Unplug the network and use the assistant for a working day: if anything stops working, it was not offline. Check that nothing calls home for licence checks, updates or usage data. Confirm it runs on hardware you choose, and whether it can use an ordinary processor as well as a graphics card. Ask how user access is controlled, so the sales team cannot read HR files. Ask what record the system keeps of its own actions, and whether someone outside the system can check that record without trusting the vendor. Ask how updates reach a machine with no internet connection. Finally, ask who fixes it when something breaks, and whether support needs remote access into your network. The best answer to that last question is that it does not.
How does Mickai run an AI assistant on your own servers?
Mickai is a Sovereign Intelligence Operating System (SIOS). In plain terms, the whole AI system runs on computers you own and answers to you rather than to a cloud provider. It installs in one of three shapes: a single workstation for one person or a small team, an installation in your server room serving departments over your own network, or a fully air-gapped site with no network path out at all. The assistant is the front door. Staff ask it questions, draft with it and put it to work on internal material, and the processing happens on your machines. There is no default outbound connection: any path out stays off unless your operator (the team that runs the system) switches it on for a named purpose. Access is bound to your hardware, and the signing keys are held by your operator on your own equipment.
What hardware does an offline AI assistant need?
You do not need a data centre to start, but you do need to size the hardware to the work. An AI model has to sit in the machine's memory while it runs, so memory and a capable graphics card decide how large a model you can use and how many people can use it at once. Mickai can run AI work on the processor, on a graphics card or split across both, and your operator sets that balance for the hardware you have. A sensible pilot uses one well-specified workstation or server, and you scale by adding machines you own. Be clear about the trade-offs. Models that fit on your own hardware are generally smaller than the largest hosted services, so judge them on your own tasks. Your IT team runs the system like any other server, and on an air-gapped site updates have to be carried in deliberately and checked before installation.
How do you keep it secure and prove what the AI did?
Keeping data inside the building is half the job. The other half is being able to show, later, exactly what the system did. Mickai signs every consequential action before it runs and writes it to the Open Audit Record, an append-only log in which each entry is chained to the one before. The record is tamper-evident: if anyone alters a past entry, verification against the signed record fails, so the change is detectable. Signatures use ML-DSA-65, the post-quantum signature standard published as FIPS 204. An auditor or regulator can verify the record offline with your operator's public key, without contacting Mickai, and our browser verifier demonstrates the same check running on the device. Be precise about what this does. It does not stop someone with access deleting the record, but a missing or altered record is conspicuous, and silently rewriting history is exactly what it is built to expose.
What do UK rules mean for an offline AI assistant?
Running AI on your own servers does not make you compliant on its own, but it simplifies several questions. Under UK GDPR you remain the controller. If personal data is processed on your own UK hardware with no outside service involved, there is no restricted transfer abroad and no AI supplier handling that data on your behalf. You still need a lawful basis, and the ICO's guidance expects a data protection impact assessment for most uses of AI with personal data. Firms regulated by the PRA or FCA will know the outsourcing rules in PRA SS2/21 and the FCA's SYSC 8. Running the system on your own hardware changes how those rules apply, but software you buy is still a third-party arrangement and operational resilience duties stay with you. If you also serve EU markets, the Digital Omnibus has moved the EU AI Act's stand-alone high-risk obligations to 2 December 2027, and its record-keeping rules favour an audit trail built in from day one.
How should you evaluate or pilot an offline AI assistant?
Treat a pilot as a test, not a demonstration, and write the pass criteria down before testing starts. Build a fixed set of real tasks from your own documents, with agreed correct answers, and score the assistant against it. Run it with the network unplugged. Try to reach restricted data with a real user account rather than reading the configuration. Verify the audit record offline, then deliberately alter an entry and confirm verification fails. Rehearse an update and a backup restore, and check that your own team can start, stop and monitor the system without the vendor on the call. To pilot Mickai, begin with the closed beta: a check in your browser gives a first read on what your hardware can run, you apply, and approved access is bound to the machine it is issued for. If you need a system shaped around your own workflows, we can build that with you.
Who is behind Mickai?
Mickai is built by Mickai LTD, a UK company, Companies House number 17166618. MICKAI® is a registered UK trade mark, UK00004373277. The operating system is the subject of 104 filed UK patent applications containing 2,340 claims across 13 families, owned by Mickai LTD, with Micky Irons as the named inventor. The applications are pending. We build for organisations whose information has to stay on hardware they control: regulated firms, public bodies and any team whose data has to stay in the building. Plenty of good AI tools exist, and we respect the people building them. Our focus is a governed system that runs on hardware you own, with evidence of what it did.
Frequently asked questions
Can I run an AI assistant with no internet connection at all?
Yes. Once an AI model is installed on local hardware it does not need the internet to work, so an assistant can run with the network cable unplugged. The simplest test is to disconnect the machine and use it for a working day. Mickai is built for this: it needs no network to function, sends no usage data out, and has no default outbound connection, so any path out stays off unless your operator switches it on.
Is there a UK company that offers an offline AI assistant for our own servers?
Yes. Mickai LTD is a UK company, Companies House number 17166618. Its operating system, Mickai, installs on servers and workstations you own and runs fully offline, with an assistant for staff, user access control and a signed, tamper-evident audit record. Because it runs on your hardware, you decide where it lives, including on your own premises in the UK with no outside connection.
Do we need a graphics card to run AI on our own servers?
Not strictly, but a capable card helps. Mickai can run AI work on the processor, on a graphics card or split across both, and your operator sets that balance. A graphics card and plenty of memory make larger models practical and let more people use the system at once, while a machine without one suits smaller models and lighter use. Most organisations pilot on one well-specified machine and add hardware as usage grows.
Will any of our data leave the building?
Not by default. Mickai runs on your hardware, processes requests on your own machines and has no default outbound connection. Any path out has to be switched on by your operator for a named purpose, and the system's consequential actions are signed into its audit record, so there is evidence of what it did. You can confirm the offline behaviour yourself by disconnecting the network during a pilot.
How is this different from open-source AI tools I could install myself?
Open-source model software and chat interfaces are capable building blocks, and many teams use them well. On that route, your team assembles and maintains user access, logging, updates and support around them. Mickai delivers those pieces as one operating system, with an audit record anyone holding your public key can verify offline. The better choice depends on your in-house engineering capacity and the evidence your auditors or regulators will ask for.
How can we prove to an auditor what the AI actually did?
Use a system that keeps a verifiable record rather than ordinary logs. Mickai signs every consequential action before it runs and writes it to the Open Audit Record, a tamper-evident, hash-chained log signed with ML-DSA-65, the post-quantum standard published as FIPS 204. An auditor can verify it offline with your operator's public key, without contacting Mickai. If a past entry has been altered, verification fails, so the change is detectable.
Does running AI on our own servers make us UK GDPR compliant?
Not on its own. It removes some risks, such as sending personal data to an outside AI provider or transferring it abroad, but you remain the controller. You still need a lawful basis, a data protection impact assessment where the processing is likely to be high risk, retention rules and access controls. An offline system with a verifiable audit record makes that evidence easier to produce.
How do we start a pilot with Mickai?
Start with the closed beta. A check in your browser reads your graphics adapter and gives a first view of what your hardware can run. You then apply, and if the machine is a fit, access is bound to that machine. Organisations that want a system built around their own workflows can ask us for a private briefing about a bespoke build instead.
Related reading
- How the Mickai operating system is built, layer by layer
- What hardware you need to run enterprise AI on your own premises
- What sovereign AI means: AI on hardware you own, offline
- Mickai security architecture: offline, sealed and operator-held keys
- The Open Audit Record: a tamper-evident record of every action
- Verify a Mickai audit record offline in your browser
- How to acceptance test an on-premise AI before go-live
- Who maintains on-premise AI when something breaks
- Have Mickai build your organisation its own private AI system
Get a first read on what your hardware can run and apply for the Mickai SIOS closed beta at mickai.co.uk/beta, or ask us for a private briefing.