What can I help with?
Private chat inside a secure enclave. Nothing you type ever leaves it.
OpenAI-compatible API
This server exposes the same models over an OpenAI-compatible API. Point any OpenAI SDK at:
Streaming and non-streaming. Also GET /v1/models, GET /warmup?model=….
Model names match the picker. If the server config sets api_key, send it as
Authorization: Bearer <key>. Requests run in the same enclave as this chat.
Enclave attestation
This app runs inside a measured confidential VM. The hardware signs a report over its own launch measurement and over the TLS key that terminates inside the enclave. Those are the numbers below, straight out of the quote.
Take a photo
The photo is attached to your next message. Nothing is recorded, and the picture stays in this browser until you send it.
Let answers use the web?
A model answering from memory alone gets details wrong and sounds certain doing it: wrong dates, wrong names, a plot point from the wrong season. With web search on, the answer is written from pages fetched at the time you ask, with sources you can open. It is the single biggest difference in answer quality here.
What it costs you, exactly
Your question, or a short query written from it, goes to a search provider outside this enclave. They see that query and this server's IP address. They do not see who you are, and they never see the rest of the conversation: the request leaves from the enclave, not from your browser.
Everything else stays as it was. With the switch off, nothing about your questions leaves this enclave at all.
Get paid to rent out your GPU
eyesoff.ai runs on hardware people rent to the fleet, and every GPU enclave is taken right now. A machine with a TEE-capable CPU and a confidential-computing GPU could be earning instead of sitting idle.
What your box would earn
- Paid by the second, in USDC on Base. The ledger escrows the runner's share of what each deployment pays and sweeps it to your own wallet for every second you hold the lease. Nobody holds your earnings for you.
- You set the price. Name your hourly rate for the box and for the card, and work paying less is refused. A full GPU card rents for about $6 an hour on this fleet today.
- No application, no approval. Your box proves what it is with a hardware attestation and the ledger does the rest. Tenants cannot read the work you host, and you cannot either.
Linux with KVM, a few GB of RAM for the guest, and a little Base ETH for gas. No public IP needed, so a machine behind CGNAT works. Setup is one command line tool; the full requirements are on the host page.
Answer quality tracks model size: small models are quick and frequently wrong. Chats stay in this browser.