Sovereign cybersecurity, within everyone's reachSovereign security for allSOC AI on your servers · free edition Request a demoDemo

Download

The free edition, on your servers

You request access, a person reads your request, and we open your account. No automated link and no bot: that is on purpose.

  • ✓ What you receive : the usage licence, the installation documentation, and the solution with its model.
  • ✓ What the machine needs : Docker, and a standard CPU is enough to start. A GPU speeds up processing markedly.
  • ✓ The licence is read before you ask: it is published in full, and it is what you accept.
  • ✓ So is the installation guide: the free edition's guide is online, step by step, with what you should see at each one.
  • ✓ The account is for downloading, not for running: the installed solution does not report to us, and your firewall counts it at zero.
◈ 100% sovereign A person answers, often within two business days.

Request access

Your information is used only to reply to you and to send you the licence. It is shared with no third party.

This information is used to review your request and open your account. A person reads it. Privacy.

The engine

You can plug in your own model

The model server is addressed by a URL. If you already run one, you point to it and the solution uses it: your procurement rules stay yours, and the model still runs on your machines.

Our figures cover the model we ship

The model supplied is fine-tuned for the solution's prompt, and that pair is what we measure. With another model, the published figures no longer apply: they are to be measured again in your environment.

  • What does not change : normalisation, correlation, the timeline, the anti-injection guardrail and the backtested tuning are deterministic. They do not depend on the model.
  • What changes : the model's verdict, its accuracy and its operating point.
  • The guard : at startup, the solution compares the model served against the one declared. On a proven contradiction it refuses to start, rather than returning verdicts quietly degraded.

Measured on 26 September 2026, by changing the prompt of a fine-tuned model: 90 % accuracy with its own, 17 % with another. The model and prompt pair is a validity condition, not a setting. We do not transpose that figure to another model: we only say that our measurements do not cover it.

Already have an account? Your access.

The installation steps are described on the deployment page, and the boundary between the two editions on the editions page.

This site uses cookies. Technical cookies are required for the site to work and stay on at all times. Functional and audience-measurement cookies wait for your consent.

Customise