Sentra is preemptive exposure AI that runs on the analyst's machine: it maps the estate, decides what has to change, and drives the tools you already run to remediate it. The engine — Elasticsearch and a fleet of playbook runners — lives in a local container, and personal data is censored before anything reaches a model.

Two passes, on-device, before the prompt goes anywhere: a local NER finds personal data across 35+ classes and 24 languages, and your own Graph directory matches the real people — names, mails, accounts — exactly. You see what was found, grouped by type and with the exact offsets, and decide entity by entity what gets censored. Answers are restored locally, so you still read real values; the model only ever saw placeholders.

The EDR and the RMM you already run — verdicts, isolations, scripts and tickets — plus your local Elastic indices (machines, accounts, sign-ins, exposures) and Microsoft Graph for the directory and sign-in telemetry. More connectors arrive as other tools are wired in.

The engine, the indices and the credentials stay on the machine, isolated per client. What leaves is the prompt, already censored, to the provider your policy allows — or nothing at all when you run a local model.

Local models through llama.cpp or MLX, or frontier APIs such as GPT and Claude. Everything outgoing passes the same censoring layer first, and the choice is per workspace.

Install Sentra, connect your EDR and RMM credentials, point it at the tenant whose directory you want behind the privacy layer, and ask the first question. The directory and fleet views fill themselves as the engine syncs.