I hate manually refreshing marketplaces looking for a deal on used hardware. So I built PriceClaw: an agentic deal hunter that watches multiple platforms every 15 minutes and only alerts me when something is genuinely worth buying.
What it does
Scrapes 5 sources every 15 minutes: Amazon, Apple Refurbished, FB Marketplace, Kijiji, eBay sold prices, and Groupchat buy/sell groups. Each listing gets parsed, SKU-classified by RAM/storage config, and stored in SQLite.
Scores every listing against a retail price matrix by config. The scorer calculates the actual discount percentage for the exact spec combination — not just “is this cheap” but “is this cheap for what it is.”
LLM triage before any alert fires. When a listing crosses the sniper threshold, it goes through GPT with five possible verdicts:
- BUY — legit deal, alert immediately
- APPROACHING — within 10% of target, worth watching
- WATCH — fair price, not urgent, no alert
- TYPO — suspiciously cheap, might be a pricing mistake, alert with caveat
- SCAM — obvious red flags, ignore
Only BUY and TYPO trigger a notification. “TYPO” is my favourite — listings at 40% of retail that might be a mistake, but worth a fast look.
iMessage alerts land on my phone with the listing link, price context, and the model’s one-sentence reasoning.
The tricky parts
TLS fingerprinting: Amazon and Cloudflare-protected retailers block standard HTTP clients. PriceClaw uses curl_cffi with Chrome 131 impersonation for those — transparent to the rest of the pipeline.
iMessage from Docker: iMessage only works via macOS AppleScript, but the workers run in Linux containers. There’s a tiny HTTP relay server on the Mac host — Docker POSTs to it via host.docker.internal, and the relay runs osascript to push the message through Messages.app.
SKU extraction from messy seller text: “M4 Pro w/ 24GB/512 — barely used” needs to map to a canonical config before scoring. A regex classifier handles 95% of cases; LLM verify handles the rest.
Delta-only price history: The SQLite price store only writes a row when the price changes — like git commits for prices. Keeps the DB compact even with continuous scraping across hundreds of listings.
Demoed at Vancouver AI Meetup
Presented PriceClaw at Vancouver AI Meetup (VAM!) in May 2026 as part of the Agentic Personal Assistants workshop. Walked away with a certificate.
Tech Stack
- Python 3.12, APScheduler, Docker Compose
- Azure OpenAI GPT-5.4 — deal triage and SKU verification
- SQLite (WAL mode) — price history, listing store, worker run logs
- curl_cffi — TLS fingerprint spoofing for hardened retailers
- FastAPI — dashboard, chat endpoint, and iMessage relay
- Playwright — FB Marketplace and WeChat scraping
- BeautifulSoup / selectolax — HTML parsing for standard retailers