An OpenAI-compatible cloud where the prompts, responses, and the model itself can be obfuscated under a key only you hold. Operator-blind by design — we compute on shapes; you read the meaning.
The compute is identical. The only thing that changes is who holds the key that gives the tokens meaning.
A stock model on stock weights. Plaintext in, plaintext out. Vanilla OpenAI clients work as-is — but anyone with the URL gets readable answers.
We transform the GGUF server-side, return the key once, and never persist it in object storage. Your client must retain that key and use EncryptedTokenizer for prompt and response handling.
You upload ciphertext the server never decrypts. The serving side needs the key out-of-band, and the client integration is not the same as plain server-managed OpenAI hosting.