How an Agent Read 471 Product Reviews in 29 Minutes with Claude Code + Oxylabs

eyishazyer · x · 2026-11-05

One Prompt, One Agent, 500+ Product Reviews

Using Claude Code (Opus 5.5) plus Oxylabs Web API's MCP tools, the author turned Meta Ray-Ban Display's public Best Buy reviews into one spreadsheet: 471 valid reviews in 29 minutes, plus 9 Reddit posts.

The Workflow

Setup: install the oxylabs web-api-mcp via uv tool install, run npx skills add oxylabs/web-api-skills, add your API key — the agent gets search and scrape tools. Cursor and Codex work too.
Prompt 1: state the job only — find public review pages, skip anything behind a login, last 12 months, no reviewer names, dedupe, save to reviews.csv.
What the agent did: 12 US-locked searches to locate review pages; 57 scrape calls paging through Best Buy listings and Reddit; 25 of those used AI extraction to pull only needed fields as JSON. 548 reviews in, 462 kept.
Prompt 2 (no new web calls): read reviews.csv, group complaints into 5-8 themes, output a table with counts, %, and 2 exact quotes with links. Four minutes later: limited apps 23%, weight and fit 15%, battery 12%.

Why It Works

Why Oxylabs Web API: country-level search, pages returned as Markdown the agent reads directly, field-named JSON extraction, JS rendering for heavy pages, and failed requests aren't charged.
Verifiability: every quote links back to its review page, so you can check the agent instead of trusting it.
The agent also flagged on its own that every Best Buy review carried the same badge: "incentivized" — a bias signal for the data.

Original post →

How an Agent Read 471 Product Reviews in 29 Minutes with Claude Code + Oxylabs

More from coding & agent