Loops HouseMetabolic AI

NEAR HACKATHON
Links
Team
1 member- ANOwner
Anshit Raj
Overview
Problem: an AI service has a cost and a price on every request, and a human watches the gap, swapping models and limits by hand when money gets tight.
Product: Metabolic AI is an AI business that runs itself; the human only sets the constraints. It sells inference: customers prepay in NEAR, each request is priced before it runs, and a deterministic policy engine (plain code, not an LLM) reads the business's own treasury and picks the best model it can afford, including open models on NEAR AI Cloud. When margins shrink it downgrades, protects its reserve and recovers on its own, logging every sale, cost and decision with a written reason on a public ledger. Agents get Spend Passes: a budget, per-request cap, allowed tiers and expiry, so an agent can spend without holding the keys. Works with Claude Code, Cursor and others via MCP and an OpenAI-compatible API. The operator can pause it at any time.
How NEAR fits: deposits are credited only after the server reads the transaction over NEAR RPC at final finality; sign-in is NEP-413 through wallet-selector; our Rust contract (metabolic-d0e41b6dc5.testnet) takes deposits and stores receipt hashes, committed with a scoped function-call key (one method, one receiver, capped gas), never a full-access key; anyone can recompute a receipt hash at /verify and check it on chain.
Scope: testnet prototype. The Stress Lab cost spike is simulated and labelled, built: NEAR Intents, Chain Signatures, mainnet, on-chain payment of providers.
Try it: https://metabolic-ai-five.vercel.app (open /demo, then /verify)