Private Gateway
Keep your inference. Make it private. Put any OpenAI-compatible base URL behind an enclave: your own model server, or a router you already use. The upstream stops seeing who you are, and never sees your keys, wallets or personal details.
What happens to a call
Who you are is dropped
Your address and headers stop at OffRouter. In the enclave, the request loses its user, metadata and caching fields. Your upstream sees the gateway calling, not you.
Your secrets become placeholders
Seed phrases, private keys, API keys, emails, card numbers, IBANs, wallet addresses, IPs and phone numbers are swapped for placeholders such as [EMAIL_1] before the call leaves the enclave. The model is told to keep them as written.
The answer comes back whole
The enclave puts the real values back in the answer, in streams and in tool call arguments too. Your code sees the same answer it always did.
Your upstream key stays sealed
It is encrypted under a key that only exists inside the measured enclave, and it is never shown again. You call the link with a key of its own, which you can throw away by deleting the link.
Every answer is signed
Each response carries a receipt: a hash of what you sent, a hash of what your upstream got, and what was redacted, signed by a key the hardware report vouches for. The preview below shows the exact request behind that second hash.
What it does not hide
The rest of your prompt
Redaction finds things with a shape: patterns and checksums. Names, places and anything else written in plain words reach your upstream as you wrote them. Your upstream still reads your prompt; it just does not get your secrets or your identity.
Images and files
Only text is redacted. An image, audio clip or file in a request is passed on as it was sent, and the receipt counts how many there were.
What your upstream does with it
Your upstream runs wherever it runs. If you want the model itself sealed too, use OffRouter's own models, which run in GPU enclaves.