One collects Google's results.
One reads the sources.
google-browser-scraper, from the proxy provider NodeMaven, is an MIT-licensed Python tool that collects Google results pages through a real Chrome browser and your own residential proxy, as a CLI, library, local HTTP API or MCP server. Levirge Search is a hosted MCP server that searches the web and reads the pages it finds. On 27 Oct 2026 we ran the same 5 queries we used for the DeepSeek comparison through both.
Same 5 queries
| Query | google-browser-scraper #1 | Levirge #1 |
|---|---|---|
| DeepSeek V4 release | deepseek.com › news › deepseek-v4-1-flash | deepseek.com/en/news/deepseek-v4-1-flash/ |
| deno compile cross-platform static binary | docs.deno.com › Runtime | docs.deno.com/runtime/reference/cli/compile/ |
| NATS JetStream KV watch example | docs.nats.io › learn › key-value › watching | natsbyexample.com/examples/kv/intro/go/ (KV intro, not watch-specific) |
| Sydney apartment rent prices 2026 | realestate.com.au › ... › Rent News | realestate.com.au/news/sydney-median-rent-hits-record-800-a-week-… |
| tokio vs smol async runtime | "Goodbye Async-Std, Welcome Smol", Substack | rustify.rs/articles/rust-async-runtimes-tokio-vs-async-std-2026 (async-std, not smol: off-target) |
| google-browser-scraper | Levirge Search | |
|---|---|---|
| Results per query | 7–9 organic, plus People Also Ask (4 on two queries) and 6–8 related searches | 10 (limit: 10) |
| Result URLs | link was null on all 41 results: 0 resolved, 41 failed. Each kept Google's /goto?url=… redirect and a display string such as docs.deno.com › Runtime | A plain URL on every result |
| Run time | 299 s for the run; 1.0–2.1 s per results page once warm | No warm-up step |
| Traffic on your connection | 66 MB, of which 59.1 MB was the warm-up | Only the MCP responses; searching runs on Levirge's servers |
Both served all 5 queries. On Google ranking, google-browser-scraper was the better list: its NATS and tokio/smol top results match the query more closely than ours. Without resolved links, though, an agent cannot open them without another step.
Where google-browser-scraper is ahead
| Capability | google-browser-scraper | Levirge Search |
|---|---|---|
| Google results page detail | organic_results with position, ads, related_questions, related_searches, pagination | title, url and content per result |
| Country | --country picks the proxy exit | Not selectable |
| Self-hosting | MIT, pip install or a Docker image | Hosted only |
| Cost visibility | Bytes and cost per 1,000 pages in every run summary | Flat plan |
| Volume | --workers N, up to 300 queries per exit | Research use; bulk scraping not allowed |
What only Levirge Search does
| Capability | google-browser-scraper | Levirge Search |
|---|---|---|
| Reading the result pages | No: it returns the results page only | fetch returns pages as Markdown, JSON and XML as raw bodies, with select and schema |
| Setup | A residential proxy with sticky sessions, plus Chrome on your machine | Add https://search.levirge.com/mcp, sign in with OAuth |
| First query | A fresh exit is warmed up first: 299 s and 59.1 MB on our run | No warm-up |
| MCP | Local stdio server, one tool: google_search(query, pages) | Remote server, nine tools, any MCP client |
| Concurrency over HTTP | One request at a time on one held identity | Search and open fetch unlimited; 1–3 full-browser slots by plan |
Where google-browser-scraper fits
SEO rank tracking and SERP monitoring: you need Google's own positions, ads and People Also Ask, by country, at volume, and you already pay for residential proxies. Its failures are labelled (captcha, block) rather than returned as empty results.
Where Levirge fits
Agents that research: find sources, then read them in full, through a browser when a site refuses plain requests, with no proxy to run.
Method. Run on 27 Oct 2026 from one machine. google-browser-scraper 0.0.6.2 (README at commit c9a7cb0; it marks the project "Early") with default settings and --no-proxy, so every query used that machine's own connection. With the residential proxy it is built for, results and link resolution may differ. Levirge through its MCP search tool, limit: 10. Results move day to day — run the same queries yourself.