useOmnis
Delegate financial tasks to AI with budgets, approvals, paid services, and proof.
Demo
About
AI agents can already research, code and operate software, but financial tasks still break the delegation loop. The moment an agent needs to buy a service or move stablecoins, the user has to return to choose providers, manage wallets, enforce budgets, approve payments and verify what actually happened. Giving the model unrestricted wallet access is not an acceptable alternative. useOmnis is a bounded financial execution agent for builders, founders and operators who already delegate work to AI and use stablecoins. A user describes the outcome in natural language, gives Omnis a budget and rules, and Omnis carries the task from intent to proof. It can understand the request, determine what capability is needed, evaluate a bounded set of paid machine services, recommend one, purchase an allowed service, stop at human approval boundaries, complete the final payment and preserve the evidence. In our flagship flow, Omnis spends $0.003 through x402 on Hedera Testnet to check a recipient wallet before a controlled 0.10 USDC settlement on Arc Testnet. The model can interpret, compare and recommend, but it cannot authorize spend, sign transactions or declare financial outcomes.
How it's made
useOmnis is built with Next.js and TypeScript around a strict separation between intelligence and financial authority. Groq’s GPT-OSS 20B handles conversational interpretation through strict JSON-schema structured output. Instead of allowing model text to become financial truth, useOmnis promotes extracted intent only after deterministic evidence validation checks the recipient, payment amount, asset, service budget, task state and approval requirements. Contextual follow-ups are slot-aware, so instructions such as “make the payment 0.3” or “use five cents for the check” update only the intended financial field. For machine-service selection, deterministic discovery first produces a bounded candidate set containing only trusted registry metadata and precomputed eligibility facts. Groq can compare those candidates and recommend one, but the recommendation is verified again against capability, availability, network policy, catalog status, exact price, remaining budget and task/registry freshness. Unknown, stale, catalog-only or over-budget recommendations fail closed. Model failure also falls back safely to deterministic selection. Recommendation output has zero execution authority. A deterministic P2 policy engine remains the spend authority. It enforces service budgets, allowed capabilities and networks, execution readiness and approval boundaries immediately before money can move. The user still explicitly triggers the service purchase. Hedera powers the machine-commerce leg: useOmnis purchases a real wallet-activity service for $0.003 through x402 on Hedera Testnet using HTS USDC. Privy provides authentication and the embedded EVM execution wallet, preserving the user-controlled approval boundary. After the service result is returned and the user approves the final action, 0.10 USDC settles on Arc Testnet. The system also maintains owner-scoped task history containing conversation state, policy snapshots, paid service evidence, approvals, settlement state and final proof. Historical tasks reopen read-only, and recommendation state is invalidated whenever task, budget, policy, registry or service facts change. Automated Playwright and unit tests cover semantic extraction, authority-bypass attempts, recommendation safety, archive isolation, execution gates and financial-write boundaries.