Inspiration

If your business bills through PayPal, someone has to send the invoices, chase the late ones, answer disputes and pay contractors, and that someone is usually the owner. I wanted to hand that job to someone you can talk to on a call, who still asks before money leaves the account.

What it does

Sign in, connect your PayPal app and press Call Penny. She joins a video call with a face and a voice and screen-shares the money picture: what came in this month, what's outstanding and which invoices are overdue. Ask her to send or chase an invoice, answer a dispute, refund a customer or pay a contractor, and she does it through your own PayPal account.

Paying out follows one rule. Anything at or under your approval threshold runs right away. Anything above it becomes a proposal, and Penny reads the amount and recipient back and waits for a spoken yes on the call. You can also approve it on the dashboard. Proposals nobody answers expire after 30 minutes.

She plans ahead too. The cash forecast dates each open invoice by how late that customer usually pays, and dragging a payment on the timeline redraws your balance. Saying "pay Sam the rest when Acme pays" sets up a payout that waits for Acme's invoice to be paid and then asks you before it goes.

When something happens on the account, like an invoice getting paid, a dispute opening or a payout failing, a PayPal webhook raises an alert and Penny asks you to take a call about it. She remembers decisions and vendor terms from one call to the next, in memory that belongs to your business alone.

How I built it

Penny is a Next.js app on Cloudflare Workers (through OpenNext) with D1 for storage. Every PayPal action goes through one PaymentProvider interface, backed by Invoicing v2, Disputes v1, Payouts v1, Payments v2 refunds, Transaction Search v1, Balances and signature-verified webhooks. Each business connects its own PayPal app and picks sandbox or live in Settings.

The call runs on LlamaCalls, which supplies the voice, the face and the LLM that decides which tools to call. Penny starts each call with an inline MCP server carrying that business's token, so the agent can only reach that one account. The worker serves 20 MCP tools covering the summary, cash forecast, invoices, reminders, disputes, payments, pay-when-paid, refunds, approvals, transactions and memory. MemoryHQ pulls facts out of calls and webhook events and gives Penny a context block when each call starts.

Reports are an AG Studio dashboard over live PayPal data, with an OpenAI-backed assistant that adds charts when you ask. The transactions page uses AG Grid, and the forecast timeline is a Bryntum Scheduler.

AI tools used

LlamaCalls runs the video call: Penny's voice, face, screen-share and the LLM that chooses which PayPal tools to call over MCP. MemoryHQ extracts lasting facts from calls and PayPal events and hands Penny a memory block at the start of each call. OpenAI (gpt-5.4-mini) powers the assistant inside the AG Studio reports page, which builds charts from live PayPal data on request.

Challenges I ran into

Keeping businesses apart took the most care. PayPal secrets and MCP tokens are sealed with AES-GCM before they reach the database. Webhooks arrive on a URL per business and are verified against that business's webhook id, and memory is scoped to one business at a time.

Retries were the other one. Approved payments use the proposal id as PayPal's idempotency key, so a retried request can't pay twice.

What's next

The payment rail sits behind one interface, so adding another provider is one new file plus a case in the provider switch. Tools, approvals, pages and calls stay as they are.

Try it

Go to https://penny.keronlabs.com and sign in with email or Google. In Settings, paste a PayPal sandbox app's client ID and secret (with Invoicing, Payouts, Transaction Search and Disputes enabled), then press Call Penny. The README covers running it yourself.

LlamaCalls and MemoryHQ are my existing products. Penny uses them as hosted services through their public APIs, and everything in this repository was written during the submission window.

Built With

  • ag-grid
  • ag-studio
  • better-auth
  • bryntum
  • cloudflare-d1
  • cloudflare-workers
  • llamacalls
  • mcp
  • memoryhq
  • next.js
  • openai
  • opennext
  • paypal
  • typescript
Share this project:

Updates

Submission history