Keep secrets out of your prompts.
Hush is a browser extension that catches personal data in your AI chat prompts before they leave your machine. It works with ChatGPT, Claude, Gemini and Grok, natively in English, Polish and Japanese.
People paste real work into AI chats every day: contracts, HR cases, client emails, bank statements. It is the fastest way to get help, and also the fastest way to hand personal data to a third party. Deleting the chat afterwards does not un‑send it. Hush exists so the AI can help with your text without ever seeing the names and numbers inside it.
You write your prompt
You type or paste text as usual, real names and numbers included.
Hush checks it before sending
When you hit send, Hush scans the text, shows you every piece of personal data it found, and swaps each one for a neutral placeholder.
The AI sees placeholders only
The provider receives your question with the placeholders in it. It can still do the work, because the structure of the text is intact.
The answer comes back complete
When the AI replies using the placeholders, Hush swaps the real values back in, right on the page.
The example uses made‑up data. The table linking placeholders to real values is stored only in your browser. It is never sent anywhere, to us or to anyone else.
A small window shows what was found. You decide what gets masked, every time.
Placeholders stay stable within a conversation, and answers are restored automatically.
English, Polish and Japanese, including checksummed national IDs: PESEL, NIP, IBAN, My Number and more.
Optional bundles for legal, HR and finance work: court case numbers, salaries, bank accounts.
The extension makes zero third‑party requests. No analytics, no update pings, no sign‑up.
The core detection runs entirely inside the extension and needs no server at all.
Extension only
Install it and it works. Detects structured identifiers offline: national IDs, bank accounts, credit cards, phone numbers, emails. Nothing to configure, no server needed.
With the local engine
An optional engine running on your own machine adds detection of names, addresses and organizations. If the engine is off, you choose the fallback: keep going with offline detection, or block sending entirely.
Hush Desktop
A one‑click desktop app with the engine built in, so the upgrade needs no technical setup at all. In development.
The local engine requires Docker and a one‑time download of language models. It listens only on your own machine (127.0.0.1), keeps no data, writes no logs of your text, and makes no calls to the internet.
A masking tool must earn more trust than the chat itself. By design, Hush reads everything you type into an AI chat. That is a serious position of trust, and no privacy policy can substitute for proof. So the entire codebase is open source under Apache 2.0: the extension, the detection rules, the local engine. Anyone can read every line, build it themselves, and confirm that the mapping between placeholders and real values never leaves the browser.
We would put it more strongly. If you cannot verify what a masking tool does with your text, adding it makes things worse, not better: you have taken the data you wanted to protect and shown it to one more party. In that situation, trusting your AI provider directly might be the safer choice. Verifiability is not a nice extra here. It is the entire argument.
Where your data lives. Placeholder mappings are stored in browser storage on your device, scoped per conversation, with expiry and one‑click erase. The extension makes no third‑party network requests of any kind. The optional engine binds to 127.0.0.1 only, is stateless, and is protected by a token so other local programs cannot use it.
Failing loudly, not silently. AI chat interfaces change often. When Hush cannot attach to a page, it shows a persistent warning that your text is not being checked, instead of quietly doing nothing. We consider saying this plainly a feature.
A seatbelt, not a wall.
Hush reduces the chance of accidentally pasting personal data into a chat. It does not guarantee that no data leaves your machine, and it can fail to intercept. Paths it does not cover:
- File uploads: an attached PDF or spreadsheet is never scanned.
- Screenshots and images: data inside pixels is invisible to Hush.
- Voice input: dictation reaches the provider directly.
- Desktop and mobile apps: Hush only works in the browser.
- Interface changes: if a provider updates its page, Hush shows a visible warning until fixed, but a warning is not protection.
- Detection misses: recognition is probabilistic, and unusual formats or uncommon names can slip through.
If any of these matter for your situation, read the full threat model before relying on Hush.