Control that follows the actual action.

Our servers do not need your work content.

Chats, documents, full logs, Skills and secrets stay in the local workspace. Identity, seats, licensing and signed policies may be handled centrally.

Every action passes the same gate.

The policy engine fails closed, organisation rules set the ceiling, and a Skill may only use the actions it declares.

• ✓Low-risk actions can be allowed and logged.
• ✓Medium-risk actions follow the organisation’s approval rules.
• ✓Payments, final sends and other red actions always require a person.

Approve what is actually about to happen.

Approval cards show the real recipient, file, command or payload — not only the model’s description. The user can approve, reject or open the full log.

Secrets stay references, not text.

API keys and connector tokens live in the operating system’s protected vault. The UI never receives their values, and an additional scrubber protects outbound data.

Untrusted content cannot rewrite the rules.

Web pages, email and files may contain instructions an AI should not follow. Omnia treats external content as potentially untrusted and can raise the risk level; the policy gate still applies.

Launch roadmap

The compliance pack is still being completed.

Local PII detection, pseudonymisation, the AI Act scanner and the final legal documentation are under development. Omnia is a technical control system, not legal advice.