Earned a Ship Weeks badge.
Post-Quantum AI Agents
Overwinter
Inference an agent pays for per call, from yield, with post-quantum keys
About
Overwinter is an inference service with an OpenAI-compatible endpoint that an agent pays for one call at a time. No account, no API key, no subscription: the endpoint answers 402 with a price, the agent pays with an ML-DSA-87 signed payment through PQ-402, and the completion comes back sealed to the agent with ML-KEM-1024. Why: an agent that depends on a monthly bill stops when someone forgets to pay it. Overwinter can be paid straight from the yield of a Perennial position: deposit once, and what the deposit earns buys the agent's calls. At 5% a year, 500 fUSD funds about one small call an hour and the capital is never spent. The agent's key can spend only that yield, so a stolen agent key cannot reach the deposit, and a classical key recovered by a quantum computer can move nothing. It also removes a practical obstacle: the agent pays in fUSD on QoreChain, and Overwinter pays the model providers however they ask to be paid.
How it's made
Plan. A TypeScript service exposing /v1/chat/completions. Unpaid requests get a 402 carrying PQ-402 payment requirements priced per call. The PQ-402 server package verifies the ML-DSA-87 payment, a relayer settles it on the QoreChain testnet (from the payer's fUSD or as a signed spend from a Perennial vault position), the request is forwarded to an upstream model, and the answer is sealed to the payer's registered QoreSeal key. An agent client built on the PQ-402 client package holds its own ML-DSA-87 key, pays, and opens the answer; it needs no gas and no EVM window. A reference agent runs on a schedule from a funded position to show the whole loop: deposit, yield, paid call, sealed answer, on-chain record. Built after PQ-402; this text will be updated as each part lands.
Timeline
Earned a Ship Weeks badge.
Earned a Ship Weeks badge.
New project.
Comments and questions
- No comments yet.