Use the same models as always – Claude, ChatGPT, Gemini and Mistral AI – for letters, research, summaries. Tegmaguard only steps in right before anything is sent: detected names, addresses and IBANs never leave the computer in the first place, only consistent stand-in data goes out. One extra safeguard before sending – or work with a locally downloaded model via Ollama, and no AI request leaves the computer.
Every feature free for 30 days. After that, viewing cases and local AI (Ollama) stay permanently free. Export (PDF, DOCX, print, email draft), deletion, Company Brain, letterhead and the cloud connection (Claude, ChatGPT, Gemini, Mistral AI) then need a license. Copying to the clipboard and Markdown export stay free.
77 seconds: what Tegmaguard does, so that detected client and customer details stay on your computer.
A quick prompt on the side is tempting and saves time – just not always in a protected setting. Depending on the content, this can raise legal questions, for instance under professional-secrecy rules or the GDPR. A Tegmaguard seat per employee steps in right there, without slowing the work down: detected names and identifiers are replaced before sending, and if a known real value (four characters or more) still slips through, the guard stops the send.
Enterprise cloud AI platforms solve a different problem: team management, one contract for many employees, EU hosting. There, contracts govern what happens to real names on someone else's servers. Tegmaguard steps in one step earlier: detected names and identifiers are replaced on your computer before the text reaches any server, including a European one. That fits professional secrecy in a law firm, practice or advisory just as well as banks, insurers or HR departments that don't want real names on someone else's servers in the first place.
| Criterion | Tegmaguard | Cloud AI platform for teams | Redacting by hand |
|---|---|---|---|
| Where does real client or customer data end up? | On your device: detected details are replaced with mirror data before sending | On the provider's server, possibly EU-hosted | At the model, if a spot gets missed |
| Professional secrecy | A technical step before sending: replacement, preview, guard | The plain text usually reaches the platform provider | Depends on how carefully the person redacting works |
| Contract and third countries (DPA, e.g. USA) | Your contract with the AI provider stays in place, but the provider only receives text in which detected details have been replaced. Real detected identities stay on your computer. With a locally downloaded model via Ollama, no external AI provider is involved | Contracts (DPA, where applicable standard contractual clauses or the Data Privacy Framework) govern what happens to the plain text at the provider and its sub-processors | Depends on what gets missed during redaction |
| Billing | License per seat, annual or monthly; AI usage through your own API key, billed directly by the provider; no usage limit inside the tool | Depends on the provider, often a subscription per user | No tool, but plenty of staff time |
| Model quality | Full strength of the connected cloud models (Claude, ChatGPT, Gemini, Mistral AI); usable because masking happens locally first | Also top-tier models, but real data travels to their servers | Any model can be used; it sees redacted spots instead of readable details |
| Setup effort | Download, ready in a few minutes, no server, no hosting cost | Depending on the provider, contract, onboarding and user management for the whole team | Hosting, servers, IT config if self-run; otherwise ongoing manual effort |
| Works offline | Yes: with Ollama and a locally downloaded model, the AI runs on the device | Usually not: SaaS, no internet means no access | Yes, but without AI assistance |
| Understands a whole case file? | Deadlines, parties, completeness, mandatory source citations | Knowledge base / RAG over uploaded documents | No |
| If a provider suffers a data breach | The AI provider holds texts in which detected details have been replaced. Details that were not detected may still be included. | Plain text held by the platform provider may be affected | Whatever was missed during redaction is affected |
| Proving deletion | Deletion record from the app (“deletion certificate”) for the data deleted in the app, with verification scan | Deletion request to the provider, proof often manual | — |
General comparison of typical approaches, not a test of specific products. The statements about cloud platforms describe the usual case and may differ for individual providers. General information, not legal advice.
Swipe sideways to keep reading
Mirror data is the default. Tegmaguard can also set values as a neutral placeholder or merely add noise. Try the difference yourself.
Automatic industry rotation paused. Compare the modes at your own pace.
"What happens in Tegmaguard, stays in Tegmaguard."
With a locally loaded model via Ollama this holds literally. With cloud AI, only text with replaced details leaves your computer; real detected identities stay with you.
Example: a payment reminder to a late-paying client, written with a local model.
The text mentioning Mr. Brandt, date of birth, street and invoice number stays on the device, on the left. On the right, Tegmaguard shows the marked-up original and, below it, what would actually be sent.
"Cologne" wasn't marked. One click opens the picker: name, title, company, address, number/ID, industry term, and whether the rule applies just to this case or permanently.
A similar substitute, a neutral placeholder, or a fixed value like "Berlin." The rule applies immediately, no restart needed.
The table lists what was replaced by what, and where the rule came from. The same person keeps the same mirror identity throughout the case.
The green note confirms: exactly this text left the device. The reply comes back with mirror data and is translated back locally.
One click on "Document" applies logo, sender and subject line, never sent to the model, only applied locally into the reply afterwards.
PDF, DOCX, Markdown, plain text, print dialog, or straight to an email draft. The file name is checked with the same detection as the content before saving.
With logo, case reference and creation date. What the model never saw is still correct in the document.
Language models love to cite statutes nobody mentioned. Tegmaguard compares the draft against your input and the case file: citations, dates and IDs without a traceable source are flagged, never adopted silently.
Works with the provider you already use, or with no cloud at all:
Claude, ChatGPT, Gemini and Mistral AI work with your own API key (cloud mode). Llama, Gemma, Qwen, DeepSeek, Phi and Mistral models run through Ollama, locally on your computer (local mode), with no provider account and, once the model is downloaded, no internet. Each AI provider's own terms apply, for Gemini those of Google. Please review them before use.
Other tools load documents into a knowledge base and answer questions from it. Tegmaguard goes further: every model-generated claim is refused unless its source resolves to a document in the case. Deterministic findings always carry their source.
Every model-generated claim — summary, timeline, answer from the case file — is refused unless its source resolves to a document in the case. Deterministic findings always carry their source: the document and, for PDFs, the page.
Client, opposing party, patient, policyholder, with role. An HMAC collision check flags when a client is listed as the opposing party elsewhere, without ever revealing the name.
Invoice, notice, medical report, reminder letter, detected by rule, with inheritance from industry-specific templates (TOML blueprints).
Dated events with stated precision instead of a guessed day. Where the case type defines deadline rules (for example tax appeals and medical files), deadlines are calculated by rules, not by a language model, and can be exported as an overview (PDF/Word).
Besides "complete" and "incomplete," the check knows "can't be verified." So you see exactly where the case file doesn't support a conclusion.
Typed comparison between documents from the same period. A finding is always a fact with a source, never a model opinion.
Standard power-of-attorney text, boilerplate phrasing, the structure of a review note. That kind of knowledge doesn't belong to a single case. It belongs to the whole firm, practice or organisation.
.tgbrain file and importable onto another computer, with its own password.
Case numbers, parties, deadlines. Draft briefs and deadline questions answered from the case file.
Diagnoses, findings, insurance numbers. Doctor's letters and case summaries, fully offline if you want.
Tax references, amounts, assessments. Review notes and invoice checks against VAT requirements.
IBANs, balances, customer data. Statement reviews and credit-decision notes.
Policies, claims, reports. Claim files with completeness checks, document comparison and claim chronology.
Personnel files, contracts, business correspondence. Where no professional-secrecy law applies, but the data still isn't anyone else's business.
All four are built into the code; only the preview can be switched off in the settings. Designed to prevent real confidential data from reaching external AI services.
A name becomes a different name, an IBAN becomes another one with a valid checksum. The model works with a consistent text.
The safety net for busy moments: before every send, a second, independent step checks the finished prompt against the known real values in the case (four characters or more). If one of them shows up, nothing gets sent, even if it was overlooked in the preview.
Only the text whose preview you've actually seen gets sent, secured with a hash, not with click order. On by default.
PBKDF2-HMAC-SHA256 with 600,000 iterations, content encrypted with Fernet (AES-128-CBC plus HMAC). Writes are atomic.
For a law firm, practice or bank, for data protection officers and IT leads, "we protect your data" is often not enough: what's asked for is evidence. Two documents Tegmaguard generates for every case at the push of a button, retrievable again at any time, as part of the annual license after the 30-day trial:
As PDF or Word, with or without the plaintext mapping: which data categories were detected and replaced, which AI provider was involved. Not a one-time thing at purchase, but generated fresh for every single case, any time: a document you can present during an audit. Can serve as an annex to your data protection impact assessment under GDPR Art. 35: the report documents the pseudonymisation before sending, the data flow and the measures taken. It does not replace your own.
Issued at every case deletion, with a verification scan that checks the deletion in the app instead of just claiming it. A document you can put in the file, not just a confirmation click.
€690 net plus VAT for businesses
Billed annually in advance. The subscription renews automatically unless you cancel. Cancel any time for the end of the current term, by email to info@tegmaguard.com or via “Cancel contracts here”. For 12 months you pay at most as much as for 10 months of monthly billing, so at least 2 months are free. All profiles (legal, health, banking) included, no extra cost.
Net: €690.00 per seat, €690.00 total, plus VAT
Prices include 19% German VAT. Businesses with a valid VAT ID from another EU country pay the net price (reverse charge). For customers outside the EU a different tax rate may apply; the final price is shown on the payment page before you buy.
Seller: Adnan Dikmen, sole proprietor (Tegmaguard), Berlin. Offer to businesses only (Section 14 BGB). Subscription per seat, monthly or annual, billed in advance, renews automatically. Cancel any time for the end of the current term. Licence file by e-mail immediately after payment. Liability: unlimited for intent and gross negligence; otherwise limited as set out in the terms. German law applies.
| Seats | Year, net | Year incl. 19% VAT | Month, net | Month incl. 19% VAT |
|---|---|---|---|---|
| 1 | €690.00 | €821.10 | €69.00 | €82.11 |
| 2–4 | €620.00 | €737.80 | €62.00 | €73.78 |
| 5–9 | €550.00 | €654.50 | €55.00 | €65.45 |
| 10–24 | €480.00 | €571.20 | €48.00 | €57.12 |
| 25–49 | €415.00 | €493.85 | €42.00 | €49.98 |
| 50 and more | €350.00 | €416.50 | €35.00 | €41.65 |
From 50 seats, the price in the row “50 and more” applies. Online orders are possible up to 49 seats; larger orders on request: Get in touch. 30-day free trial, no credit card required — download the app directly. System requirements: macOS 11 or later (Apple Silicon), Windows 10/11 (x64).
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.
Yes, for you. Tegmaguard pseudonymises: detected names, addresses and numbers are replaced with replacement values. The link back to the real values stays in your vault on your computer. Because that link remains possible, the data stays personal data for you (Art. 4(5) GDPR, recital 26). Replacement is an extra protection step before sending.
Because EU hosting, a data processing agreement and "no training" regulate the transfer to a third party, but they don't prevent it. Professionals bound by secrecy may bring in service providers to the extent the disclosure is necessary for that (in Germany, Section 203(3) of the Criminal Code, StGB). They commit an offence if they have not ensured that the provider was bound to secrecy and the provider then discloses the secret without authorisation (Section 203(4) StGB). Lawyers and tax advisers additionally need a contract in text form, comparable protection for providers abroad and, for services for a single client matter, the client's consent (Section 43e of the Federal Lawyers' Act, BRAO; Section 62a of the Tax Advisers Act, StBerG). Tegmaguard steps in earlier: detected real names and identifiers never leave your device in the first place, and the AI provider receives mirror data. More, with sources, under Law and contracts. General information, not legal advice. This describes German law as an example. Other rules may apply in your country.
Yes. The download is signed with an Apple certificate (Developer ID) and notarised by Apple. The first time you open it, macOS only shows its usual prompt for apps downloaded from the internet; the "damaged" message does not appear.
Three layers: you see a preview before every send and can flag missing values with one click. The guard checks the finished prompt against the known real values in the case and blocks on a match. Tegmaguard learns your corrections permanently, if you want it to.
A value that matches no pattern, isn't known to the case and wasn't flagged by you can slip past detection. That's what the preview is for: when on (the default), it shows you before sending what leaves your computer.
Exclusively on your computer, in a password-protected vault. There's no Tegmaguard server for your cases, no account, no sync. In terms of content, only the prompt with replacement values leaves your device, to the AI provider you chose, and with a locally downloaded model via Ollama, not even that. The app also checks GitHub for updates at every start; this transmits your IP address, no content. For the trial and purchase, Vercel, Resend and Stripe process your details; see the privacy policy. Details and limits are on the Trust & security page.
Directly with the AI provider of your choice: for Claude, ChatGPT, Gemini or Mistral AI you use your own API key, on the provider's terms. The connection goes straight from your device to the provider; Tegmaguard does not receive your texts. The provider only receives texts in which detected real details have been replaced with mirror data. With a locally downloaded model via Ollama, the AI request stays on your computer, and no external AI provider is involved. For the trial and purchase on this website, the privacy policy applies. Context with sources: Law and contracts.
On Macs with Apple Silicon (M1 or newer) running macOS 11+, and on Windows 10/11 (64-bit) as an unsigned beta. Intel Macs and Linux are currently not supported; there is no date for it.
No. On first launch, you choose your professional profile, add an API key or a local model, and set a vault password. After that, you write like in any chat.
Every feature runs through the same check chain as the very first prompt.
Name, address, IBAN, tax ID, case number, diagnosis, policy number and more, counted across all professional profiles and countries. Active at any time is whatever fits your profile and country.
Every recognised person gets a coherent fake identity: name, address, phone, email, date of birth, consistent across the whole case.
The button has three states: Check, Send, Sending … Only the text you approved in preview via a hash handshake is ever sent.
Encrypted address book, CSV import from your existing practice-management or back-office software, auto-fill when a case starts. Master data is mirror data from the first sentence.
PDF, DOCX, Markdown, plain text, print or email draft, with logo, sender and extra fields, applied locally only after the reply arrives.
Lawyer, tax adviser, healthcare, bank, insurance, HR and a general profile, each with its own pattern set; a universal profile combines the patterns of all profiles for mixed case files.
Read PDFs, DOCX, spreadsheets (CSV, XLSX), emails (.eml, .msg), meeting transcripts (.vtt), text and images (OCR) into the case, make them locally searchable (light RAG), approve each one individually before it can inform a reply.
A photo or scan is OCR-read and returned redacted; CSV/XLSX is pseudonymised column by column, formulas stay intact. Both work without an active case.
29 reviewed workflows, for example client letter, deadline check, GDPR data-access reply, doctor's letter, statement review, employment reference, plus your own, locally reviewed skills.
Lawyer, doctor, accountant, bank, insurance, HR, general, or universal: this decides which patterns get detected.
Your own API key for Claude, ChatGPT, Gemini or Mistral AI, or a local model via Ollama, with no internet at all.
From there, you write like in any chat. Clients, documents, company knowledge and skills are all optional.
Download, unarchive, drag into Applications. 30-day full-featured trial, then viewing cases and local AI (Ollama) stay permanently free — export (PDF, DOCX, print, email draft), deletion, Company Brain, letterhead and cloud AI (Claude, ChatGPT, Gemini, Mistral AI) need an annual license after that. Copying to the clipboard and Markdown export stay free.
Checksum (SHA-256) to verify: a2c1b849785dfadc78e8b948e91ec44080a114a9e9b12d214e8ce2fed2902e68
Tegmaguard.app.tar.gz.Download the installer and run it. Not yet code-signed — Windows SmartScreen and some antivirus tools will warn on first launch. 30-day full-featured trial, then viewing cases and local AI (Ollama) stay permanently free — export (PDF, DOCX, print, email draft), deletion, Company Brain, letterhead and cloud AI (Claude, ChatGPT, Gemini, Mistral AI) need an annual license after that. Copying to the clipboard and Markdown export stay free.
Antivirus keeps blocking the installer? Get the MSI installer (Tegmaguard_0.1.0_x64_en-US.msi · 63 MB, same content, different format).
Checksums (SHA-256) to verify: .exe 5f1231a165fadae9e07c9b066f6e4f9a9b1f90a3e4e6c3fc7b33bc469fb6fbbd.msi c988b91d58c74d4ea0f075b73eee021c0e2c67b8b3fb32c9e132b450d9af3a27
Windows SmartScreen shows "Windows protected your PC" on first launch because the installer isn't code-signed yet. This is an honest warning about unfamiliar software, not an error message and not a sign of malware. Here's how to run it anyway:
Tegmaguard_0.1.0_x64-setup.exe in your Downloads folder.The SmartScreen warning can appear on first launch.
Some antivirus tools (e.g. AVG CyberCapture) automatically send new, unknown installers for cloud analysis and lock the file while that runs. If you launch the installer during that window, you may see "The system cannot access the specified device, path, or file." Same root cause as SmartScreen: the installer doesn't carry a code-signing certificate yet. Not a sign of malware.
Verify it yourself instead of trusting us: compare the SHA-256 above against the downloaded file (e.g. via PowerShell: Get-FileHash Tegmaguard_0.1.0_x64-setup.exe), or upload the file to virustotal.com and check it yourself.