Your computer reads. Our server judges blind.
Deciding what is personal needs a model that understands context: "Jordan" can be a person or a country, "positive" can be a test result or a mood. BlindHush runs that model across your computer and our server without either side seeing the other's half in plain form.
1. Read locally
A small language model (13 million parameters, about 15 MB) runs on your computer's CPU and turns each word into eight numbers that describe what the word is in its context. Fast patterns catch structured values at the same time: email addresses, phone numbers, card numbers, IBANs, order and invoice numbers, dates.
2. Encrypt
The eight numbers per word are encrypted with CKKS homomorphic encryption (128-bit security). The secret key is created on your computer and never leaves it. Up to about 1,000 words fit in one 0.52 MB message.
3. Score blind
Our server runs the policy model directly on the encrypted numbers and returns encrypted scores: not personal, identifier, secret or sensitive (health, beliefs, union membership, origin, criminal records). It cannot decrypt the input or its own answer.
4. Mask, send, restore
Your computer decrypts the scores and replaces each personal value with a placeholder such as [PERSON_1] or [EMAIL_1]. The same value always gets the same placeholder in a conversation, so the AI model can still reason about who did what. The request then goes from your computer straight to Anthropic or OpenAI with your own API key. When the answer comes back, the real values are put back in on your computer, and you choose whether tools (for example, sending an email) get them too.
Who sees what
| Who | Sees | Never sees |
|---|---|---|
| You | Everything | — |
| BlindHush server | Encrypted numbers, request size and day, your balance | Your text, the results, your name or email, your IP in our records |
| The AI model | Your message with placeholders | The real names, numbers and facts |
Where it works
- Claude Code (the CLI and the Code tab in Claude Desktop) and Codex (CLI and desktop app, with an OpenAI API key): through a local proxy, set up by the installer.
- Chat in the browser (ChatGPT, Claude.ai and others): through the browser extension, which masks the text box before you send.
Install
curl -fsSL https://blindhush.com/install.sh | shirm https://blindhush.com/install.ps1 | iexThen read the install guide or what we store.