How it works
How Kaviar protects your information
Kaviar detects your sensitive details, replaces them with made-up equivalents as you send, and restores the originals in the reply, without those details ever leaving your computer. The five steps below follow a message from your keyboard to the answer.
The journey
Five steps
your computer → Kaviar → the AI → Kaviar → your computer
- You write.In Kaviar, with the AI of your choice, the way you normally would.
- Kaviar spots what’s sensitive.The message and its attachments are analyzed on your computer, as you type.
- Each detail is replaced.Mary Davis becomes Claire Martin. Dates shift by the same offset and amounts keep their proportions, so deadlines and math stay right. The mapping between originals and made-up equivalents is kept on your computer, nowhere else.
- The AI receives the protected version.The replacements are plausible and consistent with each other, so the AI replies as it would to any other message.
- The originals are restored.In the reply, Claire Martin becomes Mary Davis again before you even read it.
Protection levels
Three protection levels.
Standard.
Everyday personal data: emails, phone numbers, addresses, IBANs and cards, official ID numbers, API keys and passwords.
Enhanced.
Everything in Standard, plus people’s names, usernames, companies and dates.
Strict.
Everything in Enhanced, plus places, GPS positions and amounts. The AI’s math stays right.
Enhanced is the default. Raise or lower it at any time, conversation by conversation.
You stay in control
Something misclassified? You decide.
In a message or a reply.
Select a word and a menu lets you protect it, or leave it as is, for this conversation.
Everywhere, with the Vault.
Add it to the Vault and it will always, or never, be protected, in every conversation and at any level.
Transparent detection.
Automatic detection covers the essentials, and when it misses a detail, one click protects it.
Proof
See exactly what the AI received.
For every message, the AI view shows word for word what was sent, and how many details were protected.
Draft the employment contract for Claire Martin, employee 51907, starting 3 March.
The log
A count of what was protected.
The shield in the message field shows how many details are protected in your message, and the log keeps a running count by type, without ever storing the details themselves.
Attachments
Your files too.
Drop in a PDF, a Word document or a spreadsheet. It is protected as you send, and you see exactly what was replaced.
Web search
Web searches too.
When the AI needs the web, the query leaves protected, and results are protected before they reach the AI. In Strict mode, some replaced details, such as a place, can make results less relevant.
FAQ
Frequently asked questions about Kaviar
What is Kaviar?
Kaviar is an app that lets you use the best AIs on the market without handing them your sensitive information. Kaviar pseudonymises it: each detail is replaced with a made-up equivalent before sending, then restored in the reply, on your computer.
Does my data go through Kaviar’s servers?
Your sensitive details, never: they’re replaced on your computer before anything is sent. With Kaviar credits, the protected message passes through our servers on its way to the AI, without being stored. With your own API keys, the protected message goes straight to the provider.
Does using Kaviar mean trusting one more intermediary with my data?
No. Redaction happens on your computer before anything is sent, so your sensitive details reach neither the AI nor Kaviar. Services that anonymise your messages on their own servers have to receive them in full first.
Which AIs can I use with Kaviar?
More than 25 models from nine providers: ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), Mistral (Mistral AI), Grok (xAI), DeepSeek, Llama (Meta), Qwen (Alibaba) and Perplexity. New models are added as soon as they are released.
Why made-up equivalents rather than placeholders?
Because a consistent message keeps its context. The AI understands what it’s about, reasons on plausible values and gives the most relevant answer possible.
What does Kaviar detect?
Standard: emails, phone numbers, addresses, IBANs, cards, official ID numbers, API keys and passwords. Enhanced adds people’s names, usernames, companies and dates. Strict adds places, GPS positions and amounts. Detection works in 7 languages, on names from every culture. Its results, category by category, are published on our benchmark.
What if Kaviar misses something?
Select it and, in one click, protect it or leave it as is for this conversation. Add it to the Vault to protect it in all of them.
Can I see what the AI received?
Yes. Every message shows exactly what was sent to the AI.
Does it work with attachments?
Yes. Files are protected as you send. You see the document with its protected details, and the version the AI received.
Do I have to pay to be protected?
No. Protection is identical on every plan, and Kaviar is free to use with your own API keys.
How is this different from a DLP tool?
A DLP tool blocks the message. Kaviar lets it through without the sensitive data, so you keep using AI.
Related