Skip to content
Klarmask

Redact for real, without sending anything.

Klarmask finds the numbers, addresses and names in a PDF, a Word document, a scan or a photo, lets you decide on each one, then removes them from the file. Not just from the screen. The document never leaves your computer.

A black rectangle placed over the text

The witness, John Smith, stated that he saw the vehicle.

Select, copy, paste: The witness, John Smith, stated that he saw the vehicle.

The same document, redacted by Klarmask

The witness,           , stated that he saw the vehicle.

Select, copy, paste: The witness, , stated that he saw the vehicle.

What the tool does

  • It detects, with a check digit

    IBANs, NIR (French social security numbers), SIREN and SIRET (French company numbers), VAT numbers, bank cards, phone numbers, email addresses, postal addresses. When a number has a check digit, it is verified: an IBAN whose calculation adds up is marked “verified”.

  • It suggests names

    A model, which also runs in your browser, suggests the names of people and companies, and groups the forms of the same person: “John Smith”, “Mr Smith”, “SMITH John” are redacted in one action.

  • You decide everything

    Each suggestion is redacted or kept, with one click or one key. A name can be redacted everywhere at once. A signature or a stamp is covered by drawing an area. As long as a suggestion is still open, nothing comes out.

  • It rebuilds the file

    Each page of a PDF is rendered as an image, the black boxes are painted onto the pixels, and the searchable text is rewritten without what is redacted. A Word document is rewritten from what the tool understands: tracked changes accepted, comments and properties removed. Nothing else is copied over.

  • It checks its own work

    Before offering you the file, the tool reads it back by a second path and searches it for each redacted item. If it finds one, it gives you nothing and says where.

  • Black box or label

    Black, or a label that stays the same throughout the document: “[Person A]”, “[Company A]”, “[IBAN 1]”. The surrounding text remains readable and searchable.

  • It checks other people's work too

    A PDF or a Word document redacted by someone else? The checker looks for text still present under black areas or black highlighting, metadata, tracked changes and earlier versions.

The proof takes three steps

  1. Cut the network

    Once the page has loaded, unplug the cable or turn off Wi-Fi: the tool keeps working, from dropping the file to saving it.

  2. Watch the Network tab

    In your browser's developer tools (F12), the “Network” tab shows what the page requests. While you work: nothing. When you save, a single line: the counter.

  3. Read the rule

    The page carries a security policy that forbids the browser from connecting to any other site. The details are on the Privacy page.

Who it is for

  • Lawyers and legal teams

    Disclosed exhibits, judgments, sample contracts: without handing the case file to an online service.

  • Human resources

    Answer an access request by redacting third parties, prepare a file.

  • Public sector

    Release an administrative document after redaction, publish as open data.

  • Journalists and NGOs

    Protect a source, check a document received already redacted.

  • DPOs

    A tool with no processor to assess: documents do not leave the computer.

What it does not do

  • Guarantee that nobody will recognise the person. A redacted document can remain recognisable through its context: a job title, a date, a place. Reread it as someone who knows the case would. See pseudonymisation or anonymisation.
  • Find every name for certain. The name model suggests them and can miss some: reread, and use “Redact everywhere” for what it did not see.
  • Handle every language equally. Detection is tuned for documents written in French: French identifiers (NIR, SIREN, SIRET, tax number), phone numbers and addresses in French format, and a name model trained on French. In other languages, fewer names may be suggested.
  • Read a poor-quality scan reliably: a misrecognised word can escape detection. Reread the image, and draw an area over anything that is not suggested.
  • Keep vector quality: pages are rendered as images (200 dots per inch), with rewritten searchable text.