Names, emails, SSNs, API keys and credentials are caught and replaced with labelled tokens — in the browser, before the text reaches ChatGPT, Claude or Gemini.
Nothing is deleted. Every value becomes a label the model can still reason about, so the answer you get is the answer you wanted. And the check itself runs on your machine — a redaction tool that uploads your prompt to scan it has not solved the problem it claims to solve.
Illustration. Every particle survives the crossing in a different form — that is the difference between a label and a black bar.
The sentence keeps its shape, so the model keeps its answer.
Illustrative sample. “Please review this contract for [PERSON] at [EMAIL]” still asks the model exactly what you meant.
Silent until there is something to flag — a tool that interrupts every prompt gets uninstalled by Friday.
Named entities, regex patterns and a local model, evaluated on your machine. No network call, and no server-side copy of the text to scan.
Each value becomes a labelled token rather than a hole in the sentence, so the model still understands who did what to whom.
A badge appears on the input only when something is flagged. Replace everything, or keep the values you meant to include.
ChatGPT, Claude, Gemini, Copilot, Perplexity — it works on the input field, so it is not tied to one vendor's roadmap.
Founding licences are one payment, for good. Everyone after launch pays monthly.
Not ready to reserve?
One email when it ships. No spam.