Agent proxy
Point Claude Code or Codex at BlindHush. Every request is masked before it leaves your computer: prompts, files, command output and tool results. Tools can still get real values when you allow it.
ANTHROPIC_BASE_URL=http://localhost:8788
Privacy layer for AI assistants and agents
Talk to any AI. Keep the personal parts.
BlindHush replaces names, contact details, card numbers and sensitive facts with placeholders before your text reaches an AI model, and puts the real values back on your own computer. The service that decides what to hide works on encrypted data, so it never reads your words either.
Hi, I'm John Smith at 25 Oxford Street, London. Order 783421 from 15 October never arrived and I paid with card 4532 1234 5678 9012. My coworker Raj had a heart attack, so please email john.smith@gmail.com.
Stays here The real values never leave your computer. BlindHush keeps the link between each placeholder and its value locally.
07 00 01 03 01 01 5e 2a a4 45 8f 44 5e c4 00 20 00 00 18 01 00 00 00 00 00 14 00 02 01 00 00 00 81 00 00 00 00 00 02 01 01 00 00 00 00 81 00 00 3f 9c 1e 77 d2 08 b4 61 0a e3 5f 92 c7 13 6d 4b 28 f1 a0 59 7e 3c d8 04 bb 16 e9 42 85 2f 71 c0 9d 63 1a f8 37 5c ae 0b d4 92 66 e1 38 7f 05 cb 24 9a 50 …
0.52 MB, 128-bit Each word becomes eight numbers, encrypted on your computer. Our server scores them without ever decrypting them, and sends back encrypted scores only you can open.
Hi, I'm [PERSON_1] at [ADDRESS_1]. Order [ORDER_ID_1] from [DATE_1] never arrived and I paid with card [CREDIT_CARD_1]. My coworker [PERSON_2] had [HEALTH_1], so please email [EMAIL_1].
Still useful The model can reason about [PERSON_1] and [EMAIL_1] and write a full answer. You read it with the real names put back.
The problem
Every prompt is a copy of your data on someone else's server.
People paste customer emails, medical notes and contracts into chat windows. Coding agents read whole folders, logs and databases, then send them to a model provider with every step. Most privacy tools for this either read your text in plain form on their own servers, or block the request and stop the work.
- Your messagenames, phone numbers
- Files the agent readsaddresses, IDs, contracts
- Command and tool outputemails, keys, card numbers
- The whole history, againevery turn, re-sent
- Sensitive facts health, beliefs, records
How it works
Your computer reads. Our server judges blind. The model gets stand-ins.
-
Read locally
A small language model on your computer turns each word into eight numbers that capture what it is in context. It runs on an ordinary laptop CPU.
Your computer -
Encrypt
Those numbers are encrypted with a key that never leaves your computer (CKKS homomorphic encryption, 128-bit).
Your computer -
Score blind
Our server runs the policy model on the encrypted numbers and returns encrypted scores: not personal, identifier, secret, or sensitive.
BlindHush server -
Mask and restore
Your computer decrypts the scores, swaps values for placeholders like [PERSON_1], sends the request on, and puts the real values back in the answer.
Your computer
| Who | Sees | Never sees |
|---|---|---|
| You | Everything, with real values | — |
| BlindHush server | Encrypted numbers, request size, timing | Your text, your words, the results, your key |
| The AI model | Your message with placeholders | The real names, numbers and facts |
Products
One privacy layer, wherever you use AI.
Browser extension
Mask any text box with one shortcut before you send, or paste text into the panel and copy the masked version. Select an answer to see the real names again.
Alt + M mask the text box
Local dashboard
See what was kept from the model, by category and by app, on your own computer. It stores counts and categories only, never the real values.
localhost:8787/dashboard
Measured, not promised
Small enough to sit in front of every request.
0.52 MBsent per message, up to about 1,000 words that need checking
~0.3 sadded per chat message on a laptop-class CPU
6.3 MBof encryption keys, once per session
0plaintext characters on our server
Measured on our test setup: client on 4–8 CPU cores, server on one GPU or CPU, local network. A 50,000-token agent step uploads about 26 MB.
Where accuracy stands today
Share of values fully hidden on our chat and agent test set (614 messages). The yellow mark is our target.
Names, emails, IDs
92.9 / 98%
Names alone
92.6 / 97%
Health, beliefs, records
70.7 / 90%
We publish these because a privacy tool should say what it misses. Structured values such as emails, phone numbers, cards, orders and dates are close to complete; short first names and indirect health statements are what we are working on.
Vision
AI should know what you need, not who you are.
Most questions people ask an AI do not depend on the real name, the real card number or the real diagnosis. They depend on the shape of the problem. We want every AI conversation and every agent to work on that shape, with the personal parts staying with the person.
-
Private by construction
Privacy that rests on a promise can be broken by a breach or a subpoena. Ours rests on encryption: the server cannot read what it never receives in plain form.
-
Useful by default
Blocking a request protects data and stops the work. Placeholders let the model keep reasoning, so people keep using AI instead of working around the rules.
-
Visible to the user
People should be able to see what was kept from the model, on their own screen, and choose when real values come back.
The company
BlindHush
BlindHush builds the privacy layer between people and AI models. We combine a small on-device language model with homomorphic encryption, so masking personal data does not require trusting another company with it.
We are starting with the people who send the most data to AI every day: developers running coding agents, and teams pasting customer and patient text into chat. Pseudonymisation of this kind is what GDPR asks for when personal data has to be processed, and we are building it to be checkable by the people who rely on it.
Early access in preparation