Inspiration

AI agents can increasingly browse, book, and buy for users, but giving an agent spending power creates a serious trust problem. Existing AI wallets either expose too much money or don’t work with normal Web2 merchants. We wanted a middle path: let users express what they want in natural language, but keep every purchase constrained and human-approved.

What it does

MandatePay turns natural-language purchase requests into enforceable payment mandates.

Example request:

"Find me the cheapest Nike running shoes under S$120."

The agent then:

  1. Extracts product, budget, quantity, and preferences.
  2. Searches for qualifying products.
  3. Selects the best qualifying option.
  4. Funds a task-specific disposable virtual card using a stablecoin rail.
  5. Prepares checkout.
  6. Waits for explicit user approval before completing the purchase.

Core security model:

One agent task → one constrained disposable payment instrument → one checkout.

The user does not give the agent unrestricted access to funds. The blockchain provides the funding/settlement rail, while the disposable card provides compatibility with existing Web2 merchants.

How we built it

Our prototype architecture includes:

  • Intent layer: LLM-powered agent for extracting purchase constraints and asking clarifying questions when needed.
  • Discovery layer: Web2 product search and best-option selection.
  • Payment layer: Stablecoin-funded disposable virtual card for merchant compatibility.
  • Approval layer: Human-in-the-loop confirmation UI.
  • Settlement layer: On-chain payment verification and funding.

Planned stack: Python, TypeScript, Next.js, LangChain/OpenAI, Solidity, Ethereum/Polygon, stablecoin rail, Playwright for checkout automation.

Challenges we faced

  • Designing a mandate model that is simple enough for an MVP but secure enough to be meaningful.
  • Bridging on-chain stablecoin funding with Web2 merchant checkout.
  • Ensuring the agent cannot spend beyond the task-specific card.
  • Keeping human approval meaningful without adding unnecessary friction.

What's next

  • Real virtual card issuer integration.
  • On-chain escrow and settlement proofs.
  • Recurring or multi-step purchase mandates.
  • Richer policy constraints for more complex user intents.

Built With

Share this project:

Updates

Submission history