A hundred people on one machine

An organization with a hundred or more people who need models in their daily work has usually solved it one of two ways. Either each team has found its own vendor, and the organization's records are now spread across several of them, or a central group has stood up a service in a cloud and spends its time explaining to auditors where the data goes. Czar Keep is the third way: four or eight sovereign AI engines in one sealed chassis, in the organization's own server room, serving every one of those people from a machine that never leaves its walls.

What the people see

Czar Keep is invisible to most of the people it serves, and that is the point. An analyst opens the internal application she already uses and asks it a question; the application calls the models on the Keep over the organization's own network, and the answer comes back. A lawyer searches the organization's own contracts and gets the passage, not a summary of a summary. An engineer asks about a system and receives the runbook the answer came from. None of them signs in to a vendor, and none of them pastes a document into a browser tab that belongs to someone else.

What they notice is that the assistant knows the organization. It answers in the organization's vocabulary, from the organization's documents, and it remembers what it was told last week. That comes from Memory and Retrieval running on the same machine as Inference, with the organization's records indexed in place.

Internal applications calling the models

The first way a Czar Keep serves people is through the applications the organization already runs. The platform team gives those applications an endpoint on the organization's own network, and the models answer from there. Open-weight models run on the Keep's engines, and the platform team chooses which ones and for what. A claims system, a service console, a research notebook, a document pipeline: each calls the Keep the way it would call any internal service, and no request crosses the perimeter.

Because the Keep is four or eight engines rather than one, many applications and many people draw from it at once. That is what makes it a Grid in one chassis: the same pool a facility would build from several Czars, in one enclosure for a room that already has racks.

Retrieval over the organization's records

The second way is retrieval. Policies, contracts, filings, research notes, tickets, correspondence, and the records the organization is obliged to keep are indexed on the Czar Keep and questioned in place. Every answer carries the passage it came from, so a person can check it. The index is bounded by the organization's identity system: a person sees what they are permitted to see and nothing else. The records do not move to be indexed. The index lives beside them.

The organization's agents

The third way is agents. In Agency Studio on a Czar Keep, the organization authors agents for its own work: a triage agent for the service desk, a drafting agent for the legal group, a reconciliation agent for finance, a literature agent for the laboratory. Each is bounded to the routes and records it may touch, approved by a named person, versioned, and recorded, and any run can be replayed. Agents on a Czar Keep prepare and propose; a named person approves, and every step is on the record the organization holds. Where a Keep serves a floor, safety stays on the robot.

Adaptation and Evaluation sit behind this. The organization can tune an open-weight model on its own records and score the result against its own past work before anything is promoted, and the adapter never leaves the machine.

What the platform team sees

For the people, a Czar Keep is an answer. For the platform team, it is a machine in a rack. The team sees one appliance running the seven Substrate services, with Watchman verifying what is running and Local Bridge holding the routes the organization declared at commissioning: identity, the record systems, ticketing, logs, source control, the data stores. Everything else is denied.

The team sees the record. Every request, every retrieval, every agent run, every approval, and every update is on a record the organization holds on its own hardware, readable by the people the organization says may read it. The team sees usage, because the record shows it, but usage is not a bill. The subscription is one line for the configured appliance, fixed for the term, and a hundred people becoming two hundred does not move it.

What never leaves the walls

The prompts, the documents, the embeddings, the memory, the adapters, the record, and the models themselves stay inside the organization's perimeter. A Czar Keep can run with no route out at all. Where the organization permits connections, they are the ones declared at commissioning and no others. Updates arrive as signed bundles, by media or by the organization's own network, and support begins only when the customer asks for it through the customer portal. Nothing starts from outside.

A hundred people on one machine is not a compromise an organization accepts because it cannot use a cloud. It is the shape the work had before there was a cloud to send it to, with the models brought to the records rather than the records sent to the models.