Buterin described the configuration in an Oct. 16 post. He said he had put zkAPI, Tor, additional cover traffic and local language-model prompt rewriting into an OpenAI-compatible daemon, then created a “paranoid mode” wrapper for his pi agent harness.

Separate Tor circuits and local prompt rewriting

The system assigns each conversation a separate Tor circuit and masks requests with random cover traffic. zkAPI is intended to prevent payments from revealing the user’s identity, while a locally operated language model rewrites prompts before they are sent. Users can provide their own local model; Buterin said he was using qwen3.8-flash-next.

Buterin runs the agent harness inside a bwrap, or bubblewrap, sandbox and is also experimenting with MicroVM isolation. He additionally wraps the harness with torsocks so its network activity passes through Tor.

“It’s getting quite usable!” Buterin said. However, he said the configuration still caused a substantial increase in latency, particularly when starting a new conversation.

He also cautioned that Tor was “far from optimal” for this use. The cover-traffic mechanism can protect against passive adversaries, but is much weaker if an AI language-model provider cooperates with a Tor guard and actively manipulates traffic timing.

Buterin said this type of privacy system would require a specialized traffic-mixing network, or mixnet. Such a network would likely need constant built-in cover traffic and a higher bandwidth multiplier.

zkAPI separates payments from service requests

Open Anonymity developed zkAPI in collaboration with the Ethereum Foundation. The protocol allows users to pay in Ethereum (ETH) for access to AI and other application programming interfaces without disclosing which onchain deposit funded use of a particular service.

A user deposits ETH, pays for actual service use and can withdraw any unspent balance onchain. Zero-knowledge proofs allow the service to verify that funds are available without receiving information identifying the user’s specific deposit.

The project offers the OA Chat web interface and a local client compatible with the OpenAI API. The client can be configured to work with Open WebUI and other compatible applications.

The configuration remains experimental. In the development branch referenced by Buterin, production proving keys are not ready and the network configurations required for the main protocol have not been deployed. The privacy features therefore are not a production-ready solution.

Buterin also said cheaper zkAPI deposit and withdrawal gas costs were coming to the Ethereum mainnet soon.

Source: Incrypted