Frequently asked questions.

A private AI gateway is a layer between your team and cloud LLMs that keeps sensitive data out of the provider's hands. Minac detects personal and confidential values locally, replaces each one with a typed placeholder before the prompt leaves your environment, sends the sanitized prompt to the model, and restores the originals in the reply — so you get frontier-model answers without handing over the underlying data.

Yes. Minac connects to leading cloud models such as ChatGPT, Claude, and Gemini, so your team keeps using the model it already prefers. Whichever you choose, the prompt is sanitized the same way before it leaves: sensitive values are replaced with placeholders, and your sensitive data never reaches the provider in its original form.

Minac is built for the whole organization, not a single team. Anyone can use it like the everyday cloud model they already reach for, across every department rather than a walled-off pilot. It's especially useful for decision-makers who want a private sounding board on sensitive or early-stage work. The ideal deployment is company-wide, so protection is the default for everyone.

Those tools detect sensitive data with a single neural model. Minac runs that same class of model — locally, on infrastructure you own — but only as one of four detection layers: your own dictionary of company and client terms, exact validators for values like emails and IDs, the neural model for names and addresses, and a context-aware pass for the informal, multilingual, and context-only cases a single model misses. It then restores your values in the reply, blocks any personal data the model invents, and keeps an audit trail you control. A single-model redactor is essentially one step inside Minac — Minac is the finished gateway around it.

No. Detection runs in layers, and every prompt passes through all of them: your own dictionary of company and client terms, exact validators for structured values like emails and IDs, a neural model for names and addresses, and a context-aware pass that catches the informal, multilingual, and context-only cases the earlier layers miss. A single model or rule set leaves gaps; the layers exist to close them — because for a privacy tool, a missed value is a leak.

The usual personal identifiers — names, emails, phone numbers, addresses, and dates — alongside structured values like national IDs, IBANs, payment cards, and network identifiers, plus account names and secrets such as API keys and connection strings. You can also add a custom dictionary of terms specific to you, such as project, client, and company names. Each detected value is replaced with a typed placeholder, so the model still knows a name is a name without seeing the original.

You can run the detection step yourself — that part is open. Minac is the finished system around it: layered detection, restoration of your values in the reply, a re-scan that blocks invented personal data, document and voice handling, an audit trail, and an admin layer, all on infrastructure you own. The detection model is one component; the gateway is the product.

It leans toward protecting it. For a privacy tool, missing a real value is worse than redacting an extra one, so detection favours coverage. In Review mode, users can also see what was detected, remove anything over-cautious, and tag extra phrases by hand before sending.

Related, but aimed differently. DLP tools mostly classify and govern data sitting in your systems. Minac works at the moment a prompt would leave for a cloud model: it detects and replaces sensitive values in that prompt, restores them in the reply, and keeps the originals inside your environment. It's built specifically for using cloud AI safely, not for cataloguing stored data.

The provider receives your full prompt, but every sensitive value — dictionary terms, values Minac detects, and anything you tag in review — is swapped for a typed placeholder first, so it never sees the originals. Minac runs on infrastructure you already own, for example a dedicated virtual machine in your own Azure tenant, and your conversations stay inside that environment — retained for purposes such as audit logging and GDPR accountability, not sent anywhere else.

Yes. Alongside standard personal data, Minac keeps a customer dictionary of terms specific to you — project names, client names, even your own company name — and redacts them before prompts reach the cloud. Beyond protecting each value, this makes it harder for a provider to link prompts together and infer who you are or what you're working on.

No. The detection models run locally, inside your environment — nothing is sent out to find sensitive data. Only after a prompt is sanitized does it go to the cloud model your team chose, such as ChatGPT, Claude, or Gemini, and even then with sensitive values already replaced by placeholders.

Yes — that's the point of typed placeholders. The model receives your full prompt with each sensitive value swapped for a labelled stand-in like [NAME] or [EMAIL], so it keeps the structure and meaning needed to answer. When the reply comes back, Minac restores your real values in place of the placeholders before your team reads it.

Auto detects and replaces sensitive values silently, so protection is the default. Review adds a step where users confirm what was detected and can tag more before sending. Off means Minac doesn't change the prompt — it goes to the cloud as-is — for conversations with nothing sensitive in them. Most teams run Auto.

Detection strength is even across major European and Nordic languages — it isn't tuned to English. German, Spanish, French, Italian, Dutch, and the Nordic languages all hold up in testing. Finnish has been tested most extensively, including informal spoken and genitive name forms, and detection tolerates typos and obfuscated spellings such as name(at)company(dot)fi, not just clean text.

Yes. Alongside typed prompts, Minac processes attachments locally before anything is sent: it reads PDFs and Office documents and scrubs the text, runs OCR on images and scanned pages and redacts sensitive regions, strips image metadata, and transcribes voice notes locally so spoken details are caught too. The same replace-and-restore flow then applies.

Minac runs on infrastructure you already own — for example a dedicated virtual machine in your own cloud tenant. Detection runs there, centrally, so your team uses it from a normal browser without installing anything or relying on each person's laptop to do the work.

Yes. A factual audit trail of what was detected and replaced stays inside your environment — 30 days by default, adjustable per customer — so compliance and security teams have accountability without sensitive values being stored anywhere external.

No. The playground is a guided walkthrough using fabricated data to show the detect–replace–restore flow. It doesn't make live cloud calls or run the real detector — it's there to illustrate the experience, not to process your data.