Keep your data where it belongs.
Your data stays on your device.
Every question to an AI reveals something about you: health, money, relationships, work. Can privacy be the default — instead of an expert skill or an enterprise feature?
You protect your data yourself — discode gives you the tool for it: sensitive data is detected right on your device and suggested for anonymization before anything leaves. You decide per data point whether it goes along — because the answer really needs it. Privacy as the default — for everyone, not an enterprise add-on for pros.
Two layers, one decision per data point
Your anonymization assistant checks every request twice, right on your device: Layer 1 detects 16 data types via fixed patterns — email, phone, IBAN, credit card and more. Layer 2, HEIMDALL, is a small AI model in your browser that recognises names, companies and places — without anything leaving your device for it. Before sending, the PII review shows you every find: yellow gets anonymized (default), grey stays in plain text. Anonymized means replaced with plausible look-alikes — not a blacked-out bar but realistic placeholders, so the answer stays coherent in context. You decide, per data point.
More details
The 16 patterns: email, phone (AT/DE/intl.), IBAN (MOD-97 verified), credit card (Luhn), tax ID, VAT ID, IP, postcode+city, street, date of birth, names, insurance and account number. ~100 tech/company names (Google, SAP, React …) are ignored to avoid false positives. HEIMDALL = Home-Data-Anonymizing-Language-Layer (GLiNER, ONNX Runtime), loaded into the browser once.
Vault & Fail-Closed
The model only sees placeholders — you see real data. The mapping lives solely in the request's memory: no log, no cache, no database. After the answer it's discarded. And if anonymization fails? Then nothing is sent. No half measures.
More details
Fail-closed: on a pipeline error the system aborts with an error (503) instead of silently forwarding unanonymized.
Learns as it goes — and stays with you.
If the assistant misses something and you anonymize it yourself, it remembers that — on this exact device. Because what it recognises and learns, we upload nowhere; it exists only with you. The honest flip side: switch devices and the assistant starts over there. No sync, no backup with us — that's not an oversight but the consequence of “your data stays on your device”.
Chinese models? You decide.
Some of the best models come from China — and some data should never end up there. That's why there's a toggle: “Include Chinese LLMs”. Off means off — the routing then excludes every model hosted in China, no matter how well it scores in benchmarks. Out of the box the toggle is off: Chinese models aren't used by default — you enable them yourself only when you want to.
Our commitments
No training on your data: Most API tiers don't use your content for training anyway — and because sensitive data is replaced before sending, it never reaches a model that could learn from it. Full control: Your data is viewable, exportable and deletable at any time — no forms, no waiting, no justification needed.