QuoteCage: Safe AI Purchasing Research

Small businesses compare supplier prices manually across multiple websites. AI can help, but an autonomous browser should not expose credentials, contact suppliers, or place orders without oversight. QuoteCage turns a purchasing request into a contained research workflow. Vultr Serverless Inference plans the task, while Playwright visits supplier pages inside disposable Docker sandboxes on a Vultr VM. The application collects page information and screenshots, attempts visual verification, and presents findings with evidence and uncertainty labels. Our live demonstration researches HP 26A toner on CDW using Vultr inference. It surfaces an important ambiguity: single-cartridge and two-pack prices differ. Rather than claiming a verified purchase total, QuoteCage highlights the need for human review. Supplier-contact requests remain drafts; denying approval leaves them unsent. Sandboxes use resource limits, a read-only filesystem, dropped capabilities, restricted network access, and no inference credentials. Containers are destroyed after execution. A separate, clearly labeled safety scenario demonstrates blocked actions and approval controls. The intended customers are small-business owners and purchasing teams. Proposed monetization includes a free tier, a $29 monthly plan, and $9 research packs; billing is not implemented. The current MVP uses Docker runc. Product matching, quantity totals, shipping calculations, and reliable visual price verification remain areas for improvement. Vultr hosts the full application and execution backend, with an additional Vercel frontend.