lexform sits in your browser and anonymises everything you write to Claude and ChatGPT. The AI never sees your data, and confidentiality is preserved.
The extension needs a lexform account (free). Create my account →
Lawyers, accountants, HR, healthcare · processed in the EU · nothing is ever stored
Sovereignty
The one service you are trying to protect yourself from cannot be the one doing the protecting. So the substitution runs on our own model — an open-source model we adapted to professional documents and host ourselves, on OVH hardware in Gravelines, France 🇪🇺. No US provider anywhere in the chain, except the AI you are talking to — and it only ever receives pseudonymised data.
Your browser
The file, in the clear
What you type and what you attach stays identifiable here, on your machine, and nowhere else.
lexform · EU-hosted 🇪🇺
Substitution · OVH Gravelines
Detection and masking run in memory on our European servers, then everything is wiped. No trace, no copy.
ChatGPT, Claude…
Placeholders, and its job
The AI reads PERSON_1, COMPANY_1 and reasons just as well. It has never seen your clients.
Firewall for Claude & ChatGPT
No portal to open, no copy-pasting, no extra tool to learn. You keep writing on claude.ai and chatgpt.com; the extension handles the rest, in the background.
1
Create your account
Thirty seconds, no card required. 8,000 credits free every month.
Create my account3
Link it to your account
One key to copy from your account and paste into the extension. Once, and never again.
4
Get on with your work
The lexform badge confirms protection is on. After that you stop thinking about it.
How it works
01
You ask your question
With the real file and the real names, exactly as you would brief a colleague.
02
Identifying details are replaced
People, companies, addresses, emails, bank details, references — each becomes a neutral placeholder: PERSON_1, COMPANY_2. Done on our own model, inside the EU.
03
The AI works on placeholders
It analyses, drafts and calculates without ever holding an identifying detail. Nothing you send can be used to train it.
04
You get the real text back
The mapping never left your machine. On your screen the real names come back; for the AI, they never existed.
Try it now
Traité en mémoire vive puis effacé. Aucun envoi externe, aucun stockage.
Measured quality
Everyone claims “accurate detection”. We publish ours, measured every morning against a corpus of one hundred legal documents, with the ground truth and the scoring script available for download. You don’t have to take our word for it — you can run it again.
Detection
97.5 %
of the details that should be masked are detected — measured on French legal documents. Dates are deliberately left readable: masking a limitation period or a length of service would make the document useless.
Corpus
100 documents
1,885 details to mask, ground truth included. Corpus, scoring rules and script available for download: run it again yourself.
Last measured
19 September 2026
The benchmark runs every morning. The trend and the per-type breakdown are public.
See the measurementPricing
Free
0 € / month
8,000 credits a month free (≈ 20 pages) — one credit per word of the document analysed, reset on the 1st
No card, no commitment
Every detection type, mapping included
Prepaid credits
0,90 € HT / 100,000 credits
A 10-page contract costs about €0.04 excl. VAT
10 million credits: €90 €45 excl. VAT · valid 12 months
No subscription · processed in memory, nothing retained
Prices exclude VAT. VAT (20% in France) is added at checkout. EU businesses providing a valid VAT number are not charged it.
Privacy by design
It comes down to one sentence: the only data that is truly safe is the data that never left. Your document is processed and wiped within seconds — never stored, never read by a human, never used to train a model.
What we never do
Write your document to a disk
Keep the mapping table
Log the contents in our records
Train an AI on your data
Sell or share anything
Set tracking cookies
For your technical team
One endpoint, a Node SDK and a Python SDK. Put it in front of your own AI calls: every response returns the exact count, so there are no surprises on the invoice.
POST https://lexform.zevra.tech/api/anonymize
-H "Authorization: Bearer <your-key>"
-F "file=@contrat.pdf"
→ substituted markdown + mapping + usage count
Frequently asked questions
The questions a firm asks before installing, answered the way we answer them on the phone.
No. The document is anonymised in France first: names, companies, addresses, IBANs and identifiers are replaced by tokens. Only those tokens cross the Atlantic. And if our engine does not answer, the message is blocked — nothing ever leaves in the clear.
On a sovereign AI we host in France, in Gravelines. Your document does not leave European soil during anonymisation. Only the already-anonymised text reaches the AI you use.
No. Content is processed in memory and erased within seconds: nothing on disk, nothing in our logs, nothing in a training set. All we keep is a word count, which billing requires.
You alone. The table mapping each token to its original value stays in your browser: we never send it anywhere and keep no copy. We would therefore be unable to re-identify your documents, even if asked to.
lexform keeps the identity of your clients and their cases out of whatever leaves your machine, which is where the risk with generative AI lies. Privilege remains your personal duty: the tool reduces exposure, the judgement stays yours.
No, and that is deliberate. In legal work dates carry the reasoning: limitation periods, seniority, notice periods, deadlines. Masking them would make the document useless to the AI — precisely what you are trying to do with it. A date of birth therefore remains visible to the model.
Twenty pages a month are free, no card required. Beyond that you top up prepaid credits: one credit is one word of the analysed document. You pay only for what you send, and the balance is returned with every call.
The Chrome extension, from the Chrome Web Store, then a token to paste once. It sits between your browser and the AI: you write as usual, and anonymisation and restoration happen without any extra step.
Strictly under the regulation, no: anonymisation there is irreversible, whereas here you can restore the real names. Technically this is pseudonymisation — and that is what you want for a case file: a document only you can unmask. We say “anonymisation” because that is the word everyone uses; the mapping table itself never leaves your machine.