Inspiration
AI inference should be accessible to people and agents. Gyblos is being built as a marketplace where buyers can use AI models and providers can offer inference. For this hackathon, we focused on making the payment experience simple: approve a spending limit once, chat, and pay for recorded usage.
What it does
In the Gyblos desktop prototype, a buyer chooses PayPal Sandbox and authorizes a session. They can send multiple prompts to an Ollama model within the approved limit. Gyblos displays the amount recorded against the limit; at the end, the buyer can review and close the session, triggering a capture of the recorded usage. The demo uses test funds and a demonstration price of $0.10 per completed chat response. The authorization is a spending limit, not a deposit into a USDC wallet.
The prototype uses one local seller with a PayPal Sandbox business account and local Ollama inference. Each chat message is currently independent. Gyblos also has a prototype local connection for coding agents, although our demonstration focuses on the desktop chat and PayPal flow.
How we built it
We built the desktop experience with Electron, React, TypeScript and Node.js. The payment flow uses PayPal's Sandbox APIs for authorization and capture, and Ollama runs the model that generates the responses. A local session record tracks usage and helps the app handle payment status and recovery.
Challenges and what we learned
A payment authorization is not the same as a completed charge. We had to connect multiple inference requests to a single approved limit, clearly show recorded usage, and handle uncertain outcomes without blindly retrying a payment. Designing the flow taught us how important payment states and reconciliation are to making an AI experience trustworthy.
What's next
We want to improve the buyer experience, expand agent integrations, and develop the marketplace toward connections with remote providers. The PayPal implementation shown here is a sandbox prototype.
