Dispatcher
The on-ramp. Routes any AI task through providers you use today, audited and budget-capped.
Encrypted context, customer-held keys. The persistence layer behind the Cowork desktop app.
Token providers are commodity. Switch in an afternoon. Years of encrypted context are a relationship. Own the memory, own the customer.
The model is open-weight. Compute is a commodity. Any sovereign provider with comparable hardware can serve the same model class. The switching cost is one afternoon and a new endpoint URL.
Your accumulated context (every conversation, every uploaded document, every embedding) lives in one place. You cannot move years of encrypted state in an afternoon. The customer holds the key. That makes the relationship.
Customer-held keys, ciphertext at rest, sealed lookup, portable by contract. The four properties that distinguish a moat from a database.
A single key derived on the customer's device. Xanadu never holds a copy. The customer can rotate, revoke, or destroy the key without operator participation.
Every document, every prior session, every embedding stored encrypted under the customer key. The storage layer has no plaintext to surrender.
Retrieval against the Memory happens inside the same TEE that runs the inference. Plaintext lives only in the enclave, only during a request, and only for the customer who holds the key.
The Memory format will be documented and the schema versioned. If you leave, you export the ciphertext blob and your key; another sovereign provider can read it without coordination from us.
The Cloud runs the inference. Memory makes the next session remember the last one. Cowork is the desktop app that surfaces both.
The on-ramp. Routes any AI task through providers you use today, audited and budget-capped.
The engine. Confidential inference in Reykjanesbær, operated by a Swiss Stiftung in formation in Zürich.
The moat. Your encrypted context layered behind the Cowork desktop app. Customer-held keys.