whoami && why
Surplus Intelligence is a decentralized exchange for LLM inference — sellers list their OpenAI-compatible endpoints, buyers get routed to the cheapest seller at request time, settlement is on-chain USDC on Base. It's a marketplace sitting on top of every provider.
surp.ivc.lol sits one layer above: we don't host models, we don't run a seller, we don't even hold your money. We watch Surplus's live order book, classify every model into a small set of useful combos (best-coding, pro-reasoning, etc.), and route your request to whichever model is cheapest for the combo you asked for — for a transparent 5% markup on the spot price.
The LLM API market is broken in a familiar way: providers charge what they think they can get away with, prices are opaque, and switching cost is high. Surplus Intelligence already solved the supply side — a real market where sellers compete on price. But on the demand side, you still have to know which of 145 models to call. That's not a product; that's homework.
Combos turn the homework into a one-word choice. surp/best-coding is "give me the cheapest model that's actually good at coding, whatever that is right now". The user doesn't need to track model releases, benchmark changes, or price moves — the gateway does that continuously, 24/7.
x402 — HTTP 402, the "payment required" status code that browsers ignored for 30 years — lets us charge per request with no account, no API key, no subscription, no KYC. You hold USDC in a wallet; you sign one EIP-3009 authorization per request; we settle on Base; the response streams back. That's the entire onboarding flow. It's the native payment rail for autonomous agents, and it's what makes a micropayment-priced service viable at all — no Stripe fees eating the margin, no minimum-topup, no metering infrastructure to maintain.
This project supersedes qb.ivc.lol (QuestBase). QuestBase was an AI scavenger-hunt protocol for crypto conferences — a fun idea, but a niche product. surp.ivc.lol is a tool we actually use ourselves, every day: the same Hermes agent that powers our infrastructure can point at surp/best-coding and get the cheapest coding LLM on the internet without thinking about it. If it's useful to us, it's probably useful to the broader community of agents, tinkerers, and builders who want cheap inference without the overhead.
The gateway code is small and readable. The routing logic — the combo resolver and the class taxonomy — is the valuable part, and it's all in one file. We'd rather the community fork it, improve the class definitions, add combos, and submit PRs than have us be the single source of truth. If you think best-coding should include a model we missed, or that we need a best-agents combo, open an issue.
this is a community project. we make it work for us first, and share it because the marginal cost of letting you use it is approximately zero.