Your privacy, itemized

What is proven, what is architecture, and what still needs your trust. We list all three, because anyone who claims everything is proven is lying.

Encrypted channel to confidential hardware

by design

Checking the live connection...

Chats live only on this device

verified live

This browser currently holds 0 chats with 0 messages in local storage. We have no database for your conversations. Clear your browser data and they are gone forever, even we could not bring them back.

AI runs inside sealed hardware

by design

Inference runs on NVIDIA H100 GPUs in confidential-computing mode inside AMD SEV-SNP encrypted VMs, operated by Privatemode (Edgeless Systems, Germany). The hardware signs a measurement of the exact code it runs; the encryption proxy rejects the connection unless that signature chain verifies against AMD and NVIDIA root certificates. Neither the operator nor the datacenter can read prompts, and nothing is retained after the response.

No account, no tracking

by design

No sign-up, no cookies for tracking, no analytics scripts on the chat. There is nothing to link your conversations to you.

This deployment matches the public code

you trust us

Deploy provenance is being set up: soon this will show the exact public commit this server was built from.

Our relay server

you trust us

Your message passes through our server on its way to the encrypted channel, and today you have to trust that we do not log it. We do not, and our code is public, but honesty requires the label. The fix is already on our roadmap: verification moves into the app on your device, so our server becomes a blind pipe. The upcoming native app closes this gap completely.

Verify this deployment yourself

Four checks, increasing effort. Any mismatch means our claims are broken; please report it.

1. Ask the server what it runs:

curl -s https://privepal.com/api/version

2. Read exactly that code: the response links to the commit on GitHub.
3. Confirm the pages you receive are stamped with the same commit:
```
curl -s https://privepal.com | grep -o 'privepal-commit[^>]*'
```

4. Rebuild the app from source and compare the served bundles:

git clone https://github.com/yachty66/Privepal
cd Privepal/web && git checkout <commit>
npm ci && npm run build
# compare .next/static/chunks with what privepal.com serves

Honest limits: these checks catch any mismatch in the code your browser runs. The server-side relay cannot be externally proven this way; a deliberately malicious operator could misreport its own version. Closing that gap requires hardware-attested serving, which is on our roadmap, and is already how the AI inference itself runs.

Don't take our word for it

Our code is open source. One click sends it to an AI you already trust, with the question: is this really as private as they claim? It's a second opinion, not a formal proof, but it's a second opinion we can't influence.

Ask Claude Ask ChatGPT Ask Perplexity

If the AI can't fetch URLs, use "Copy audit package": it puts the question plus the complete source of the privacy-critical files in your clipboard, ready to paste into any AI. Raw bundle: /audit-bundle.txt

Want to check the deep claims yourself? Privatemode's attestation flow, manifests, and source are public: docs.privatemode.ai