One cancelled workshop. A good next plan for every guest.
Try the interactive preview → · PayPal AI Hackathon entry
Watch the 143-second demo (2:23) · YouTube · A real PayPal sandbox capture and $68 refund, followed by a separate AG Studio rehearsal: two families, one last place, and recovery of a simulated refund. Live AI in both workspaces. Recording notes and transcript.
Running a cancelled class today? Download the free planning sheet, or have a plan prepared for US$29 per batch, up to 30 bookings. The assisted service delivers a worksheet, open questions and guest-message drafts; the operator executes changes in their own tools. Scope and worked example.
A kiln breaks. Four bookings need attention. One pair wants to split: a sister can attend Sunday, the other person wants her money back. Another guest wants a full refund. A third group wants to move together. Someone still needs the available times.
Raincheck turns those replies into specific rebookings and refunds, holds the remaining places, and follows every payment through to a settled outcome.
- Open the preview and select Mina.
- Click Analyze reply. Her two tickets become one Sunday place + a $68 refund.
- Open Try a recovery scenario, select the interrupted-response option, and run the plan.
- Recover the refund. The same saved operation settles one refund and one seat.
- Try Noah's full refund, Jules's two-person transfer, and Priya's open question.
The public preview uses fictional bookings, simulated payments, and recorded real AI outputs. It runs privately in your browser; Reset demo clears its local workspace. The Node app also supports live structured AI and a PayPal sandbox adapter.
Small teaching businesses publicly describe PayPal payments, limited places, and full refunds when the teacher cancels. Bookwhen's own workflow spans contacting guests, transferring tickets, refunding through the payment gateway, and recording the refund back in the booking system.
Raincheck puts a whole cancelled class on one desk: customer choices, places left, pending questions, and payment outcomes.
Research and first-party sources · How the recovery works · Live model observations
Requires Node.js 24.5+; developed with Node.js 26. The core app uses built-in HTTP, SQLite, and fetch. Core dependencies to install: 0. The optional AG Studio module has a separate SDK and licence.
git clone https://github.com/blucca/raincheck.git
cd raincheck
npm start
# http://127.0.0.1:4326The default app replays the recorded examples with a durable simulated payment ledger. Data lives in temp/runtime/; set RAINCHECK_DATA_DIR to choose a private directory.
npm testFocused checks cover cancellation outcomes, restart recovery, seat capacity, exact ticket/quote validation, and the sandbox checkout API contract.
The interpreter calls an OpenAI-compatible structured-output endpoint. The demonstrated model is NVIDIA Nemotron 3 Super 120B A12B on Nebius Token Factory.
Copy .env.example to a private environment file and configure the model values. Create two private budgets from budget.example.json: NEBIUS_BUDGET_FILE for the account-wide allowance and RAINCHECK_BUDGET_FILE for Raincheck's own cap. Supply the account's confirmed credit allowance, chosen caps, model prices, and expiry. Every model request reserves its maximum cost durably in both budgets before HTTP; missing or exhausted budgets stop the request. The sample budget starts inactive.
node --use-env-proxy --env-file=/path/to/private/raincheck.env src/server.mjsNew replies use live AI. The application computes amounts and checks availability from the saved booking. AI selects intent; the teacher confirms the proposed action.
The integration uses the free PayPal sandbox exclusively. The app now includes a three-step checkout desk:
- Configure
PAYPAL_SANDBOX_CLIENT_IDandPAYPAL_SANDBOX_CLIENT_SECRETin a private environment file. Use a freshRAINCHECK_DATA_DIR. - Open local Sandbox checkout and create Mina’s two-seat, $136 test order.
- Approve it with a virtual Personal sandbox buyer, then click Capture approved order.
- Raincheck reads the capture and matches the booking, amount, currency, and order ID. The capture is saved directly into the cancellation workspace.
- Open the cancellation desk, analyze Mina’s reply, and execute one Sunday seat + one $68 refund.
Complete sandbox setup and recovery guide →
The default sandbox workspace has one fictional guest with a payment awaiting capture. The default simulated workspace retains all four stories. Each workspace keeps its payment mode and payment history across restarts.
A captured ticket can be refunded in full. A multi-person booking can split between full ticket refunds and equal-price rebookings. PayPal-Request-Id is saved with the exact amount and capture before a refund call. Known refund IDs are reconciled by GET.
Sandbox verified — October 12, 2026 (UTC): an account-connected run completed OAuth → $136 order → virtual buyer approval → capture → $68 refund → independent GET receipt. A newly worded reply used live Nebius AI, and a server restart preserved the completed case and replacement seat with zero holds. Exact input, results, and public verification summary → Sandbox funds are virtual; the public preview uses simulated payments and recorded AI.
Two families. One Sunday place. Choose a make-up class in the native Studio filter, see its competing household plans, and open the existing confirmation desk. A custom seat-pressure widget keeps whole-batch availability beside the matching families' proposed, in-progress, and settled refunds. Refresh reads the result after a household settles.
The optional local module uses AG Studio 3.0.0 with your own valid licence. Its native Agent Framework can query the saved batch and filter the family queue through the existing budgeted model connection.
The film's Studio segment starts at 1:33. Its five fictional households have scripted plans and simulated payments. The live analyst reads the batch and filters Mina and Elliot; the confirmation desk then interrupts and resumes Mina's saved refund operation. Refresh shows Sunday at 8 booked, 0 held, $68 settled, and $136 in other proposed refunds across the batch.
Local setup and repeatable five-family example
This first version handles one merchant's captured USD tickets, fixed per-ticket prices, equal-price replacement classes, and teacher-approved outcomes. The cancellation policy gives each guest a choice of a replacement class or their full ticket value. The teacher's local workspace is the source of seat availability; external booking-system write-back is a future integration.
Run one Node process for each workspace database. Browser previews use a local fictional workspace; live server deployments need their own access and customer-data controls.
New work for the PayPal AI Hackathon, started October 12, 2026. Research, product design, code, and test work use AI agents. Runtime AI and its observed failures are documented. The MIT credit-budget module is adapted from DockProof; this product's class-recovery workflow, interface, payment adapter, and shared browser/server state machine are new.
MIT. Feedback from teachers and workshop operators is welcome through GitHub issues or email.