Privacy & Security

User (browser) ──HTTPS──▶ Web server ──HTTPS──▶ Phala TEE (Intel TDX) ──HTTPS──▶ Venice AI

│ RAM encrypted by CPU hardware │ No SSH access │ No shell │ Distroless image │

Best-price router (/router, used by the Studio): Phala TEE ──HTTPS──▶ Surplus · AntSeed · OpenRouter · Chutes · Engy · Venice (direct)

Only when a market is cheaper, or when Carpe's own network cannot serve the request. These markets process it outside the TEE, under their own policies, and some may see and keep it: a private chat or embedding model never goes to those (below). Privacy Policy, section 6.

Private chat and embedding models stay private. Venice runs a model marked Private on infrastructure it controls, without logs; for chat and embeddings, the router only sends it to a market that keeps nothing of it either: Carpe, Venice direct, Chutes, which runs in confidential compute, or Engy, on the models it serves on its own GPUs with zero data retention. Images, voice and video do not have this filter yet: there, Private only describes how Venice runs the model. Anonymized models (Claude, GPT, Gemini…) are forwarded by Venice to their maker, which sees the content but not who sent it, and keeps it under its own policy. Each model says which it is, in the Studio and in /v1/models (privacy). Through the API, for chat and embeddings, "privacy": true keeps the request in a TEE, and "privacy": false routes even a private model on price alone.