Intent store
What should be true. Allocations, records, ownership, policy. Rows and columns, ordinary SQL, ordinary backups. It answers counting and ownership questions.
"Which addresses in this subnet are still free, and who owns this range?"
Most architecture pages describe a platform. This one follows a single DHCP lease from the moment a device picks up an address to the moment the name resolves, and names every part it passes through on the way, in plain language rather than product names.
Everything on this page is a path that runs today, except where it says a part is in development.
Seven steps, in the order they actually happen. Step through them, or let it run.
Drag the diagram sideways to follow the whole path.
A laptop, a camera, a handset joins the network. Your DHCP server gives it an address, exactly as it does today.
The lease is posted to OmniTwin, signed with a key you control. No agent on the device, no polling loop waiting to notice.
The message lands in a queue drained by a pool of workers. A burst of leases is absorbed rather than lost.
The compiled math core works out the subnet and the reverse name for that address. Deterministic: the same address always gives the same answer.
The DNS records are written to the intent store, and the address appears in the reality store, connected to the things around it and tagged to your tenant.
The zone change is pushed out to the DNS providers you already use, so the name resolves where it always did.
A reconciler sweeps reverse DNS in the background and checks it against what OmniTwin holds. If it could not finish the sweep, it says so rather than claiming a clean pass.
Three columns: what you already run, what OmniTwin runs, and where the answers are kept. Moving lines are live data paths. Select any part to read what it does in one paragraph.
Drag the map sideways to see the whole system.
live data path where answers are kept drag sideways on a narrow screen
One database would be simpler to draw and worse to live with. The two questions engineers actually ask have different shapes, so they are kept in the shape that answers them.
What should be true. Allocations, records, ownership, policy. Rows and columns, ordinary SQL, ordinary backups. It answers counting and ownership questions.
"Which addresses in this subnet are still free, and who owns this range?"
What is true, and how it connects. The live picture, held as a graph. It answers relationship questions, the ones that turn into slow joins in a table and into a short walk in a graph.
"What sits behind this address, and what else goes dark if this link does?"
The gap between the two is the signal worth acting on. When what should be true and what is true stop matching, that difference is drift. Detecting drift automatically, and agents that would act on it, are in development. The architecture page goes into the engines underneath.
Alongside the guardrails below, every change follows one path before it touches production. The what-if check and the human approval both exist today.
What stops a bad change
Five stages. Four are wired. The last one, agent proposals, is planned.
It starts as a proposal, from a person today and from an agent later. Nothing is applied yet.
The change is tried against the twin in memory first, so its blast radius is known before production sees it. This ships today as ordinary code, not as an agent.
Approval is the default. Auto-approval would be opt-in, per type of change.
The change goes in and is logged, so you can see what changed and who changed it.
Changes proposed by an agent would flow through this same path. No agent is wired today.
Every table holding customer data is scoped to a tenant by the database itself. A query that forgets to filter still returns nothing from another tenant, because the rule does not rely on the query being careful.
The keys your DHCP servers sign with are re-read continually. Withdraw one and it stops working quickly, without a restart or a deploy.
Address maths runs in one compiled library, never in the browser. Two engineers looking at the same subnet see the same numbers, because only one thing computed them.
The reconciler reports a clean pass only when it finished one. Half a sweep is reported as half a sweep. Quiet rounding up is how teams end up trusting a stale picture.
Updates are pushed out to your DNS. Nothing needs a hole through your firewall to reach in, which is usually the first question a security review asks.
You can ask what a change would do before you make it. This preflight check ships today as ordinary code, not as an agent, and it changes nothing.
OmniTwin sits alongside what you run rather than in front of it. If it stops, addresses are still handed out and names still resolve.
The beta runs against your DHCP and your DNS, read-only until you say otherwise.