Purse — one authorized payment relationship, protected budgets for people and agents
One monthly budget. As many isolated Purses as you need. No cross-pocket borrowing or agent self-approval.
Purse is an experimental governance layer for AI-assisted commerce, evolving from the LunchPilot lunch-order prototype. An administrator creates a budget envelope, assigns Purses to people or constrained agents, and configures spending templates for food, supplies, shipping, and personal purchases. Users get a simple monthly-to-daily planning experience and an agent that recommends only within the applicable rules. A Red Button produces an auditable end-of-day budget review and flags exceptions.
Demonstrable sandbox capabilities
Separate Purse clusters isolate administrators, people, instruments, budgets, and records; cross-cluster access is denied. Nested monthly templates and per-order limits; idempotent simulated holds, protected identity/owner checks in the backend test harness, no automatic borrowing. A localhost demo includes meal recommendations, a fictional item catalog, simulated delivery and tracking, post-delivery ratings, and a mobile-oriented personal-budget screen. Administrator exceptions must be approved within the correct scope. Agents and other Purses cannot approve or fund them. End-of-day Red Button reviews the tracked budget ledger, exposes unresolved items, and suggests equivalent choices rather than undisclosed substitutions.
PayPal + AI integration approach
The standalone LunchPilot prototype contains a PayPal Orders v2 Sandbox create → buyer approval → capture adapter and an optional OpenAI model recommendation adapter. The server independently rejects recommendations outside the allowed items. A separate in-process contract test harness binds a spending-pocket authorization preflight to a simulated PayPal gateway and proves unauthorized requests do not invoke that gateway.
Honest current limitation: No authenticated PayPal Sandbox capture, live OpenAI API call, or fully atomic connection of the Purse pocket ledger to the active HTTP checkout route has been verified. The current demonstrations use fictional money, merchants and orders. There is no real PayPal Credit delegation, PayPal-held Purse balance, actual delivery service, or live purchasing.
Local proof status — November 6, 2026
The existing private Purse repository passed 205 automated tests locally; the corresponding sandbox PR GitHub CI completed successfully on Python 3.11 and 3.13. The required candidate Find It control suite passed five tests, but its inspected registry contained no registered Purse System ID, so production promotion remains blocked.
This project page describes the product direction and verified private sandbox work. It is not a completed qualifying PayPal AI Hackathon submission. Before formal entry: establish and record a genuine PayPal Sandbox transaction and AI inference, prepare a specifically approved public licensed and runnable competition repository, publish an accurate under-three-minute YouTube demonstration, and provide the remaining submission answers.
Why it matters
Families, students, founders, and accounting departments need spending automation that respects scope, not just checkout automation. Purse proposes reusable budget policies and independent ledgers around delegated purchasing, with the account holder retaining authority. A single-bill credit or prepaid arrangement is a product proposal requiring provider/issuer approval, not an existing PayPal feature.
Built With
- javascript
- openai-api-optional
- paypal-orders-v2-sandbox-adapter
- python
- sqlite