Budgets for anything that spends on your behalf.

AI agents are starting to buy things: API calls, data, compute, often for a fraction of a cent each. Today their owners have three bad choices: hand over a card with no real ceiling, pay on the blockchain per call (slower and costlier than the call itself), or ask sellers to trust an agent they’ve never met. The same gap exists for people: parents, employers and gift givers who want to hand over money for one place without handing over a card.

The owner sets aside a budget in USDC for one seller, one spender and one end date. The spender pays with signed slips carrying the running total. The seller checks each slip on its own machine in milliseconds and collects everything later in one transaction. The spender can’t authorize more than the budget, the money can only reach the named seller, and every valid slip is already backed by money set aside for that seller. The owner takes back what’s left after the end date.

• Agents can now use paid tools on their own (MCP, tool use), and HTTP 402 “Payment Required” is being revived as a way to charge them.
• Dollar stablecoins on low-fee networks make one settlement for many small payments practical.
• Owners need a hard spending limit that lives outside the prompt, because prompts can be hijacked.

• Wedge: paid APIs and data services that want to charge agents per call without card fees or per-call transactions.
• Second front: closed-loop prepaid at places people pay often (canteens, cafés, events, suppliers), where offline acceptance for regulars matters.
• Same contract and SDK for both, so every seller that joins serves both kinds of spender.

The contract charges no fee, has no owner and no token, and will stay that way. Revenue options sit around it:

• A hosted collector that submits sellers’ payments and pays the network fee, for a monthly price.
• A hosted, durable seller backend for APIs that don’t want to run their own payment store.
• An owner dashboard with budget requests, receipts and team controls for companies running many agents.
• Card-to-USDC top-ups through a licensed on-ramp partner, on a referral basis.

• Rebuilt from scratch starting 23 Sep 2026.
• Contract deployed and verified on Arbitrum Sepolia, with 41 Foundry tests including invariant tests at 256 runs × depth 50.
• Live test run: an agent made 20 paid calls, settled in 3 transactions, and the seller received exactly the 0.25 USDC it served. With its network cut, payments kept being accepted; a thief with the stolen key was refused three times.
• Working dashboard, shop till with offline acceptance, phone wallet, and an MCP server for Claude.
• Not yet audited. No mainnet deployment yet. No paying users yet.

Contract on Arbitrum Sepolia: 0xb9ae…67F2 · source (MIT)

Pilot partners (a paid API or data service, and one canteen or café) and funding for a security audit before a capped mainnet launch (100 USDC per budget, 1,000 USDC per deployment).

Built by raldblox. Contact through GitHub.