Sonar Desk

Market signals you only pay for when they are right, bought by an AI agent that cannot spend without you.

Comment

Inspiration

Paid trading signals are sold on trust. A paid group charges a monthly fee whether its calls are right or wrong. Sonar already publishes a public, walk-forward record of its prediction-market calls, so we wanted to put money behind it: if a call is wrong, the buyer gets refunded automatically. And because people increasingly hand research to AI agents, we wanted an agent that can buy these signals for you, inside a budget you set, without ever moving money on its own.

What it does

  • Signal Briefs, $1.99 through PayPal. Every open market where Sonar has a call sells a brief through the PayPal JS SDK v6 button. Buying freezes Sonar's call at that moment (side, score, price). An AI model then writes the brief from Sonar's own numbers only: the trade tape, the price path and other venues.
  • Accuracy guarantee. When the market resolves against the frozen call, or is cancelled, Sonar refunds the capture through the PayPal Refunds API automatically. The guarantee's record (right / refunded) is public on /desk.
  • Shopping agent. On /desk you set a mandate (budget, categories, least call strength) and chat. The agent searches Sonar's radar, opens a PayPal order for each brief that fits, and stops: every order waits for you to approve it in PayPal. The budget and limits are enforced in code on every purchase, not by the model, and a repeat purchase is refused by the server. It can then read and summarise the briefs you paid for.
  • Sonar Pro ($7.99 a month, PayPal Subscriptions): every brief included, no checkout per brief.
  • Webhooks, signature-verified: if you approve and close the tab, the webhook still captures your order and unlocks the brief; refunds and subscription changes update the record.

About the refund in the video. No market with a bought call resolves before judging (the markets that currently have a call close in 2027). So the video shows a labelled sandbox demo settlement: a secret-protected, sandbox-only route settles one brief as if its market resolved against the call. The outcome is simulated; the refund is real and runs the same guarantee code and the PayPal Refunds API. Demo settlements are listed separately on /desk and never counted in the public record.

How we built it

Next.js 14 on Render. PayPal: Orders v2 (create and capture), Refunds, Subscriptions (product, plan, subscription), Webhooks with signature verification, the JS SDK v6 one-time payment session, and the PayPal Agent Toolkit (@paypal/agent-toolkit/ai-sdk, read-only get_order) inside the agent. Every create, capture and refund carries a PayPal-Request-Id, so a retried sweep can never refund twice.

AI tools and how we used them:

  • Vercel AI SDK runs the shopping agent's tool-calling loop (find calls, open an order, check purchases, read a brief).
  • gpt-oss-120b on Groq is the model for the agent and for the brief writer. The writer gets only Sonar's own facts as JSON and must not invent numbers; a deterministic template takes over if the model is unavailable.
  • PayPal Agent Toolkit gives the agent read-only access to its orders.

Data: Sonar's existing radar of Panta prediction markets on Solana, read from the shared Upstash store. The Desk deployment runs read-only, so it can never write the radar.

What changed since 1 October

Sonar existed before the hackathon (the market radar and its track record). Everything PayPal and everything agentic is new and was built from 11 October: the brief checkout and writer, the accuracy guarantee and refund sweep, the shopping agent and its mandate gate, Sonar Pro, webhooks, the /desk pages, the sandbox demo settlement, and the read-only deployment mode. Branch paypal: about 3,700 lines across 38 files, 15 new tests (144 in total).

Challenges we ran into

  • Making an agent safe with money. The model never gets a tool that pays. It opens orders; a person approves in PayPal; our server captures only after approval. Budget and category checks live in src/lib/mandate.ts and run on every purchase.
  • Refunds must never happen twice. Idempotency keys on every PayPal write mean a retried sweep returns the first result.
  • v6 presentationMode: 'auto' fails in some tabs (Payment Request), so we use the popup and keep an approve-link fallback.
  • Groq retired the Llama model mid-build. We moved to gpt-oss-120b and made tool schemas accept null for optional fields.

Accomplishments that we're proud of

A refund guarantee that actually runs end to end in the PayPal sandbox, and an agent that shops but cannot spend.

What we learned

The safest agentic-commerce design keeps the model on the "propose" side: it can find, explain and open an order, while limits and money movement stay in code and with the person.

What's next for Sonar Desk

Partial refunds scaled to how wrong a call was; Pro members credited for missed calls; the agent running on a schedule against a standing mandate; going live with a PayPal business account.

Built With

  • groq
  • next.js
  • paypal
  • paypal-agent-toolkit
  • paypal-subscriptions
  • paypal-webhooks
  • render
  • solana
  • typescript
  • upstash
  • vercel-ai-sdk
Share this project:

Updates

Submission history