haggleface

Inspiration

We’ve all been there. You want to buy a used PS5, a camera lens, or a bike. The best prices are hidden on Facebook Marketplace or eBay, but finding them is an absolute nightmare.

Spending hours mindlessly refreshing tabs, trying to filter out scams, finding the best deal, and enduring painful, ghosting-prone negotiations with sellers just to save a few bucks. It is a massive, universal pain point.

We asked ourselves: why do enterprise companies have multi-million dollar automated procurement software to buy supplies, while everyday consumers are forced to manually browse across 15 different tabs?

That frustration birthed haggleface. We wanted to stop haggling and start automating. Our goal was to build a true autonomous agent that doesn't just show you links, but acts as a tireless, tactical proxy to hunt down the best deals, verify the quality, and handle all the awkward haggling for us.

What it does

haggleface is an autonomous procurement agent that turns the chaotic secondhand web into a unified trading desk.

• Cross-Marketplace Swarm: Concurrently browses different marketplaces (e.g. Facebook Marketplace, eBay, etc.) to find the best deals for what you need.

• Live Agent Feed: Watch the Steel.dev cloud browser physically navigate and inspect listings in real time directly in the UI.

• Smart Deal Scoring: Automatically normalizes messy listings and ranks them based on price deviation, condition, and seller quality.

• Defect Detection: Scanning listing photos to spot hidden flaws (scratches, wear) and automatically using them to justify steeper discounts.

• One-Click Negotiation: Select a listing and the AI drafts a tactical opening offer. You review and approve it before the agent sends it via the marketplace chat.

How we built it

• Frontend: Next.js, TypeScript, Tailwind CSS, and shadcn/ui for a high-density, dark-mode dashboard.
• Browser Infrastructure: Steel.dev provisioned our stealth cloud browsers, while Playwright handled DOM traversal to bypass aggressive anti-bot protections.
• Backend: Convex managed the reactive database, instantly streaming agent events and scraped data to the client.
• AI Layer: Vercel AI SDK and multimodal LLMs processed images, sanitized raw HTML, and drafted context-aware negotiation strategies.

Challenges we ran into

Web agents are tough because they have to be flexible to many obstacles, such as bad internet, captchas and weird UI. We spent many hours making sure the agent could properly navigate each platform we want to host.

In addition, major sites such as Ebay use aggressive screening for bot-like behavior, which was hard to get around for our web agents to interact with listings. It was tough navigating our agents through these IP blocks to properly find relevant deals and actually contact sellers.

We also realized that we were the only team trying to scrape marketplaces like facebook and ebay, which has no apis and has not been done before. Exploring this new frontier was an exciting challenge for sure.

Turning unstructured data into a clean dashboard was tough. Early iterations were cluttered and boring. We iterated heavily to build a sleek, terminal-inspired interface that balances the data grid with the live browser feed.

What's next for haggleface

Broader Marketplace Swarms: Expanding beyond Facebook and eBay to include local and niche platforms like Karrot, OfferUp, and Craigslist.

Safe-Pickup Logistics & Scheduling: Parsing chat agreements to coordinate meetup times, automatically suggesting verified safe-exchange spots (like local police stations or busy transit hubs), and adding the pickup to your calendar.

Seller Fraud & Flake Shield: Cross-referencing seller profiles across platforms to spot serial flakers, suspicious pricing anomalies, and newly created burner accounts before dispatching an offer.

Built With

• clerk
• convex
• next.js
• playwright
• react
• steel.dev
• tailwind
• typescript
• vercel
• zod