edilsonsantos-co-za/owed

An AI collections copilot for freelancers and small studios who bill through PayPal invoices.
Late payments are the cash-flow problem of small service businesses: the work is done, the invoice is
sent, and then someone has to notice it is overdue, work out how hard to push this particular client, and
write the awkward email. Most freelancers either chase too late or chase everyone the same way.
Owed reads your PayPal invoices, learns how each client actually pays, and has Claude agents draft the right
reminder for each overdue invoice: friendly for a good client who slipped, firm for a habitual late payer,
final for 45+ days. Every draft lands in an approval queue; nothing reaches a client until you press
Approve & send, which sends the reminder through PayPal's Invoicing API. All of it sits inside an
AG Studio dashboard the same agents can rebuild on request.
Built for the PayPal AI Hackathon (Oct–Nov 2026).
## What it does
Live receivables dashboard (AG Studio) over your PayPal sandbox invoices: outstanding and overdue,
aging by client, open invoices most-overdue-first, and a cash forecast that shifts each due date by
that client's real payment habit.
Client payment profiles: average days to pay, days late, on-time rate and a 0–100 risk score per client,
computed from PayPal invoice and payment history.
A team of agents in the dashboard chat: AG Studio's five built-in agents (lead, planning, data, page,
widget) plus Owed's own collections agent. Ask "who owes me money? draft reminders" and the lead
delegates to collections; ask "add a chart of overdue by client" and it delegates to the dashboard agents.
Human-in-the-loop by design: agents can only draft. You edit the subject or message, then approve;
Owed calls POST /v2/invoicing/invoices/{id}/remind (or creates and sends a new invoice). An audit log
records who did what: agent drafted, you sent.
Morning brief: a server-side Claude agent uses PayPal's Agent Toolkit (read-only tools:
listinvoices, getinvoice, listtransactions, listdisputes…) to investigate the account and
write a three-line briefing with today's focus clients.
Demo mode: with no PayPal credentials, Owed runs on a built-in scenario (seven clients, six months of
invoices) so anyone can try it in under a minute.
## How AI is used
| Where | Model / tool | What it does |
| Dashboard chat (6 agents) | Claude Opus 5.5 (Sonnet 5.5 selectable) via the Anthropic Messages API | AG Studio's Agent Framework runs the agents in the browser; Owed's custom Claude adapter (lib/studio/claude-adapter.ts → /api/ai/turn → lib/claude-turn.ts) translates each turn to streaming Claude calls with adaptive thinking, and Claude's stream back into AG-UI events. |
| Collections agent | Claude + Owed tools (getreceivables, getclientprofile, draftreminder, draftinvoice) + Studio's viewschema/executequery | Judges tone from history, writes reminders and invoices into the approval queue. |
| Morning brief | Claude + @paypal/agent-toolkit (Bedrock-shaped tools mapped to Claude tool use) | Agentic, read-only investigation of the PayPal account; structured JSON output. |
Server-side refusal fallback is enabled on Claude calls (fallbacks: "default"), and the hosted demo caps AI