Agent PayGuard
A pre-signing control prototype for AI-agent purchases, using Intercepta recipient checks.
About
Agent PayGuard explores how an AI agent can buy a digital service within the limits set by its owner. The demo purchase is a structure report for a bundled Solidity contract. The seller returns an x402 payment quote, and the buyer checks the quoted recipient before any payment can proceed. Intercepta supplies address evidence; the application's own policy determines how that evidence should affect the purchase. The control boundary brings together the recipient, network, asset, amount, task budget and available risk evidence. The goal is to prevent a changed order, an uncertain response or a repeated request from becoming an unintended payment. The current build demonstrates a local HTTP 402 quote, a server-side Intercepta address-check integration, and clearly labelled offline policy scenarios. Signing and settlement remain disconnected. A successful paid-report purchase, an end-to-end live-risk block and an autonomous AI buyer are not yet demonstrated. Mainnet address evidence is not presented as proof of testnet coverage, and synthetic examples are not payment receipts.
How it's made
We built the prototype in TypeScript, with a React/Vite interface and separate Express buyer and seller services in one repository. The seller uses the x402 SDK to quote a report for a bundled Solidity sample. The report contains structural and lexical observations; it is not a security audit. The buyer's server-side Intercepta adapter keeps the API key out of the browser and checks the payment recipient. The repository includes guarded-signing and payment-operation modules, along with offline tests for transaction binding, spending limits, repeated requests, uncertain execution outcomes and report validation. These tests exercise local control logic with fixtures; they do not prove an onchain payment. The running application keeps signing and settlement disabled. Intercepta provides risk evidence, while our own policy and execution controls are separate responsibilities. We used Codex to assist implementation, tests, documentation and interface iteration, and OpenAI image generation for the hand-and-eye logo. The project owner defined the purchasing scenario, directed the product and interface revisions, and specified the acceptance requirements. AI-assisted source areas include apps/web, apps/buyer, apps/seller and docs. No application-level AI provider is integrated into the current purchase flow.