how it works

what happens to your data, step by step.

01

we remove sensitive data

when you press send, limenn reads your prompt and any attached files in your browser, before anything goes out. it looks for personal data in three passes, each one catching what the previous one can't:

  1. fixed formats. emails, phone numbers, IBANs, AHV numbers, card numbers, IP addresses, postcodes. identifiers with a checksum are verified, so a random number isn't mistaken for an IBAN.
  2. names in context. a language model that runs on your own computer reads the sentence the way a person would and finds people, organisations, places and addresses, in German, French and English, even in lowercase or ALL CAPS.
  3. a second look at the doubtful cases. is "Sommer" a surname or a season? only words the first two passes are unsure about get a second check, with all other personal data already masked.

every value found is replaced by a placeholder that says what it was: [PERSON-1], [EMAIL-1], [IBAN-1]. the same value always gets the same placeholder, across your prompt, your files and the whole conversation, so the AI can still tell who is who.

what limenn recognises

what should stay can stay: product names, courts, legal references and the AI's own name are left untouched, so the request still makes sense. how well each type is caught ↗

02

your AI works on the clean request

limenn sends the cleaned prompt through your normal chat, in your own account. the AI sees a complete, readable request. it just reads [PERSON-1] where you wrote a name.

language models handle this well: they reason about [PERSON-1] the way they would about "the patient" or "the tenant", and repeat the placeholders exactly in their answer. what the task needs, such as a diagnosis, a notice period or an amount, can stay in the text.

documents are cleaned the same way and attached as text. with the document vault you prepare a folder once; after that, attaching a cleaned file is instant.

03

your result comes back with the original context

when the answer arrives, limenn looks up each placeholder in your local vault and shows the real value in its place, highlighted, directly in the chat. you read and copy the answer as if the AI had known the names all along.

the vault, the list of which placeholder stands for what, lives only in your browser. it is never uploaded and is cleared when you close it.

what leaves your device, and what doesn't

stays with you

  • the names, numbers and addresses you typed
  • your original documents
  • the vault that maps placeholders to real values
  • the detection model, which runs on your computer

goes out

  • the cleaned prompt and cleaned documents, to the AI you chose, through your own account
  • for the second look only: a single doubtful word, with everything around it masked. you can switch this off
  • a one-time download of the detection model when you first install

how to install

limenn is in early access. join the waitlist and we'll send you the install link.

in your browser claude.ai · chatgpt.com

  1. add limenn to Chrome from the link in your invite. it takes one click.
  2. open claude.ai or chatgpt.com. the limenn bar appears above the chat. the first time, limenn downloads its detection model once (about 135 MB); after that it runs offline.
  3. type in the limenn bar instead of the normal box. check the highlighted values, then send.
  4. read the answer with your real data filled back in.

with Claude Desktop macOS

  1. install the limenn companion. it lives in your menu bar.
  2. click "install into Claude Desktop" in the menu, then restart Claude.
  3. press ⌘⇧A to open the limenn bar. allow limenn under Accessibility once, so it can paste the cleaned message into Claude for you.
  4. optional: choose a document folder. limenn prepares your files once, and Claude can then read the cleaned versions.

want to try it with your own documents?

join the waitlist ↗