Tegmaguard replaces personal details on your computer before a text reaches an AI provider. This page explains how that works technically, which connections the app makes and what you control yourself. It is based on the current program code and its documentation.
Tegmaguard was founded in Berlin in 2026, and your questions are answered directly by the development team. Reference customers are not yet available; the technical statements on this page are based on the program code and its automated tests.
You work with Claude, ChatGPT, Gemini or Mistral AI as usual. Before anything is sent, Tegmaguard replaces, on your computer, whatever makes a person or a case identifiable.
Replacement runs on your computer, and your texts never pass through a Tegmaguard server. These are the connections the app makes:
How this is checked: an automated test calls every program interface of the app in local mode and logs every attempt to connect to the outside. Result: none. The test does not cover the update check in the app shell, and a firewall measurement of the finished app is still outstanding.
Everything stays on your device. What the vault encrypts and what you secure yourself:
The mapping between real details and replacement values, the client register and your API keys sit in a password-protected vault. Key derived with PBKDF2-HMAC-SHA256 at 600,000 iterations, content encrypted with Fernet (AES-128-CBC plus HMAC-SHA256).
Derived replacement values, for a date of birth or an IBAN for example, are generated with a random secret that exists only in your vault. Someone who knows the replacement and the source code therefore can't recover the real value by trial and error.
Exports such as PDF or Word files are stored in plain text wherever you save them, as are the settings including letterhead details, temporary print files, logs and your own templates. For these we recommend disk encryption: FileVault (macOS) or BitLocker (Windows).
When you delete a case, the app removes it from the vault, overwrites the backup copies and issues a deletion certificate. Exports you have saved and operating-system backups such as Time Machine are yours to delete.
Four steps, defined in the program code:
Detected details are swapped for consistent replacement values: a name for another name, an IBAN for another one with a valid check digit. The model works with readable text.
When on (the default), the preview shows you before sending what leaves your computer, and you can mark anything that was missed. Exactly the approved text is sent, secured by a hash. It can only be switched off for text the app checks automatically.
Right before sending, a second, independent step checks the final text and the model instructions against the known real details of the case. If it finds one (values of four characters or more, see “Initials and everyday words”), nothing is sent. You can only release matches one value at a time.
The reply is translated back on your device. Beforehand, the app checks whether any real details appear in it and flags every match. This check is for the record; it does not hold the reply back.
Detection works with rules, name lists and context, optionally supported by a local model via Ollama. That keeps every replacement traceable; the preview shows you each one. Where rules don't reach, the preview puts the decision in your hands.
Detection is tested continuously against new sample texts; gaps that turn up are documented and fixed. We don't quote a blanket detection rate, because it would be no promise for your texts.
Here we summarise statutes and case law in general terms, each with a link to the source. This is not legal advice: how you use Tegmaguard and an AI service is your decision and your responsibility as the controller under the GDPR, together with your advisers and your data protection officer where needed. Summary as of 1 November 2026.
Tegmaguard was founded in Berlin in 2026 by Adnan Dikmen and is developed together with his team. We have been working on the app for about a year, first on planning and research, then on programming.
