← Back to the blogSovereignty

Running an AI Tool Without Your Code Leaving the Building

A hospital IT team wants AI help on its appointment module. The CISO has three questions. Here is how an on-premise, air-gapped setup answers each one.

October 29, 2026 · 4 min read

Picture a hospital IT team that maintains the module patients use to book, move and cancel consultations. The code is old and the rules are dense (who can book which specialist, how far ahead, what happens to a slot when a referral expires), and nobody on the current team wrote most of it. A tool that reads the module and explains it back would save them weeks.

Then the CISO asks three questions. Where does the code go? Who gets to see it? What stays with the vendor afterwards?

In a hospital, those questions are far from paperwork. The appointment module handles patient identities, schedules and care pathways, and the code itself describes exactly how that data moves. If the answers are vague, the project stops. It should.

The worry reaches well beyond hospitals. In a Cognizant study published in October 2025, 85% of 1,000 senior executives at Global 2000 organizations said they were concerned about their technology estate’s ability to integrate AI. The same study expects spending on modernizing legacy systems to rise from 26% to 43% of IT budgets by 2030. A lot of old code is about to be touched, and nobody wants to lose control of it in the process.

So let’s take the CISO’s questions one at a time.

“Where does the code go?”

Nowhere. Lyriks installs on your infrastructure with a single command, runs on-premise, and works air-gapped on an isolated network with no internet access. It makes no outbound connections. Nothing reaches our servers, whether code, extracted spec or metadata.

For the hospital team, that means the appointment module is analyzed on the same network that already hosts it. The security review needs no data-flow diagram, because there is no flow leaving the building to draw.

The same holds for what the tool produces. The extracted spec and the graph that stores it live on the same machines as the code. Unplug the outside connection tomorrow and nothing changes in how it runs, because it never relied on that connection in the first place.

YOUR INFRASTRUCTURELYRIKSENGINEDeterministic checksPROJECT MEMORYSpec, rules, decisions,history, all linkedYOUR SOURCESCode, tickets,docs, databasesYOUR LLMYour model,your tokensREADSPECCODEONE COMMAND · ON-PREMISE OR AIR-GAPPEDOutsideNOTHINGLEAVES
Inside your infrastructure, one Lyriks block holds both the deterministic engine and the project memory (spec, rules, decisions, history). It reads your sources, hands the spec to your own LLM and checks the code it returns. Nothing leaves.

“Who gets to see it?”

Your people, and the model you pick. That second part deserves a closer look, because many tools blur two different jobs.

One job is writing. That is your LLM, and you choose it. If your policy allows Claude Code, Cursor or Copilot, the validated spec goes to them over MCP with its rules and their sources. If it doesn’t, you plug in a local or sovereign model running on your own servers, with your own tokens.

The other job is checking, and that part uses no AI at all. The verification engine is deterministic, built on graph rewriting and developed in-house. Same question, same answer. Because checking depends on no outside service, it behaves identically whether you switch models tomorrow or stay on a closed network for years.

Take a rule from the appointment module, something like “a follow-up slot cannot be booked before the referral date.” The engine checks every change against that rule in exactly the same way, whichever model wrote the change.

“What stays with the vendor?”

Nothing, since the vendor never received anything. The more interesting question is what stays with you.

Teams that work with an AI assistant slowly build their understanding of a project inside chat histories. Change tools, or watch the model change underneath you, and that understanding evaporates. You start explaining the referral rules all over again.

Here the project memory is independent of any LLM. The follow-up rule above lives in a graph database on your side, linked to where it came from: the scheduling service, the appointments table, the ticket that introduced it. Swap models and the rule, its sources and its links stay put. Models come and go, and the memory stays. You can also export it in full, traceability included, so you are not locked in with us either.

YOUR INFRASTRUCTURELYRIKS · PROJECT MEMORYScheduling serviceAppointments tableRULENo follow-up before referralThe ticket behind itMODELS, SWAPPABLEMCPModel todayMCPModel tomorrowSWAPTHE VENDORReceives nothingEXPORTIn full,traceability includedmodels come and go, the memory stays with you
The follow-up rule of the appointment module lives in the Lyriks project memory on your side, linked to the scheduling service, the appointments table and its ticket. Models plug in over MCP and can be swapped, the vendor receives nothing, and the memory exports in full with its traceability.

Getting it past the review

A tool can be sovereign on paper and still take nine months to approve. Three things shorten that.

The install is one command, so there is no integration project between you and a first result. The code is open source under the AGPL, so your security team can audit it before deployment and confirm there are no outbound calls. And the scope can stay small. The hospital team could start with the appointment module alone, judge the extracted spec and its links to the code, and read the compliance report on a first change before anyone mentions the rest of the information system.

Lyriks also owns its IP outright, with no dependency on a foreign platform vendor. Some procurement teams care a great deal about that, others not at all, and you already know which kind yours is.

The Security page covers the deployment architecture and network isolation options in more detail, written for infrastructure teams and CISOs. If you would rather walk through a setup close to yours, book a demo.

Go further

Security pack · free · 14 pages

Lyriks security pack, for CISOs

Deployment model, network flows, data and the LLM question, plus pre-filled answers to the usual security questionnaire.

Keep your software in tune

Start with one module. Lyriks recovers its spec, then verifies every change. All it needs is your repository.