All use cases
All professions

Summarize an email thread or meeting notes

Paste an email thread. The widget masks names and addresses as you type. Claude returns a summary you read in clear text on the human side only.

The problem

Nobody plans to leak a client's name. It happens because summarizing a forty-message thread by hand is miserable, the chat box is right there, and the paste takes two seconds. No policy survives that gap. The fix has to sit in the moment of pasting, not in a training module.

How it works

  1. Paste as usual

    The widget sits on top of Claude, ChatGPT or Cursor. It reads what you are about to send, not what you have already sent.

  2. Replace in one click

    Detected values are offered with their protected equivalent. Accept and the draft is rewritten before it is sent; the model receives the marker.

  3. Read the answer in clear

    When the reply comes back carrying markers, the real values are revealed in place, over the text. You read plain language; the model never held it.

Prompt to Claude

Summarize this thread between [a.PERSON_1] and [a.PERSON_2] about the [c.DEAL_X] project.

What gets masked here

  • Sender and recipient names
  • Email addresses and phone numbers
  • Company and project names
  • Meeting locations and dates
  • Any value you mark as sensitive yourself

What you get back

The convenience of pasting into a chat box, without the exposure that normally comes with it, and no change to the tool your team already opens every morning.

How accurate is the detection?

Questions on this use case

Which AI tools does the widget work with?
It is an overlay rather than an integration, so it sits on top of Claude, ChatGPT, Cursor and the other assistants your team already uses. Nothing has to be rebuilt around it and no vendor has to grant access, which is what makes it work across tools that change quickly.
What if it misses something?
You see every detection before sending and can add your own. That matters because no detector is perfect: on 5,000 multilingual samples Hexagone AI detected 96.95% of ground-truth personal data against 75.30% for Microsoft Presidio, which is a large margin and still not a guarantee. The review step is what closes the gap.

Try it on this exact task.

One week free, no credit card. Nothing is uploaded, so you can run it on a real file without asking anyone's permission.

macOS Apple Silicon and Windows 10/11