THE WORKSPACE TRUST MODEL LOCAL PII INTELLIGENCE, EXTERNAL REASONING
Customer-controlled environment
Workspace client — web & desktop

Chat · files · images · audio · generated files. Sign-in via your identity provider (OIDC/SAML SSO).

Self-hosted PII model + Salus Engine

Every message is masked in-house before any provider call — real values become PHONE_8f3a. Vault, policy, restore, and audit run here.

tokens → ← response
Outside world
Delegated external models

Each task routes to a suitable model — code, reasoning, research, documents — with automatic fallback. All of them see tokens only.

OpenAI Claude Gemini Azure Mistral
and any OpenAI-compatible provider

A client designed around protection#

The Gateway protects other people's applications; the Workspace is the application. Because Salus owns the client, protection is native rather than intercepted: masking, human review, restore, and policy are part of the product experience, not a layer bolted on around someone else's UI. Users get a modern AI workspace — chat, file uploads, images, audio, generated output — and the organization gets the same guarantees as everywhere else in Salus.

How a message travels#

The self-hosted PII model — Salus's own detection intelligence running on your infrastructure — masks every message before any provider call. The app then delegates each task to an external model suited to it: code questions, long-document work, research, and general reasoning each route to a strong provider, with automatic fallback if one is unavailable. Those external models receive tokens only, and the answer is restored in-app by the same restore pipeline the Gateway uses. The Workspace is not a self-hosted answering model — external models write the answers; the self-hosted intelligence manages what they are allowed to see.

More than chat#

  • Files and multimodal content — PDFs, Office documents, images, and audio are anonymized on upload, with OCR and visual detection where configured.
  • Generated files — documents and artifacts produced by the model are restored inside your perimeter before delivery.
  • Built-in agent gateway — an OpenAI-compatible endpoint routes your own agents and applications through the same mask → delegate → restore path as the chat clients.
  • Model routing — the per-task delegation described above, applied automatically so users never choose a provider to get a good answer.

Enterprise fit#

Sign-in federates to your identity provider via OIDC/SAML — MFA and conditional access are inherited, roles map from group claims, and no passwords are stored in Salus. Governance is shared with the Gateway: the same Salus Engine, the same vault, one policy plane, one token-only audit stream. Two products, one engine — running both is the normal pattern, and Gateway vs. Workspace explains how they divide the work.