Camoufox
The highest ceiling we measured — if you pay the babysitting tax.
Camoufox is a Firefox-based anti-detect browser. It's the one tool we tested in three configurations, and the spread between them is the real finding.
Stock headless (0.4.11): 4 of 21 — Target, Rightmove, Reddit, Wikipedia. That already includes Reddit, which every headless Chromium variant failed. Headed on a real display, same version: 3 of 21, no stealth gain, plus new flakiness (Target itself timed out with a transport error). Simply showing a window bought nothing.
Tuned (cloverlabs-camoufox 0.6.0: headed, persistent browser profile, geoip, adaptive settle, humanized interaction): 11 of 21 machine-verified — and a manual review of the saved screenshots says 14 of 21 rendered real, usable content. AliExpress, Indeed, and Reddit all displayed genuine results before the adapter's 80-second timeout killed the attempt. The tuned profile uniquely passed Walmart, eBay, Booking, Airbnb, Trustpilot, CoinMarketCap, and YouTube — sites nothing else on the bench touched.
The cost side of the ledger is just as real: attempts ran 14–80 seconds, six targets died on transport timeouts, and the hard walls held — Zillow's press-and-hold verification, Realtor's block page, Yelp and Craigslist's refusals. This is a tool you operate, not a tool you call.
Test conditions
21 real public sites (public-tough-v1), 2026-05-24, no proxy, single residential IP, Linux. Deterministic validators: expected status code plus required and forbidden content markers. Pass rates measure the tool and the network position together. Without a proxy, results reflect how each site's protections treat one plain residential connection as much as the tool itself.
What worked
- +Tuned profile: 11/21 machine-passed, 14/21 visually usable — more than double anything else we tested without paid infrastructure.
- +Unique passes on Walmart, eBay, Booking, Airbnb, Trustpilot, CoinMarketCap, and YouTube.
- +Even stock headless passed Reddit, which defeated every Chromium-based tool.
What failed
- −Operational drag: 14–80 s per page, six transport-timeout deaths in the tuned run, persistent-profile state to manage.
- −Headed mode alone (no tuning) was worse than headless — 3/21 with new flakiness.
- −Hard interaction walls still win: Zillow press-and-hold, Realtor, Yelp, Craigslist.
- −Three of the tuned 'failures' actually rendered usable content before the 80 s adapter timeout — measurement itself gets hard at this depth.
When to use it
High-value hard targets where a 50–65% unproxied success rate justifies slow, stateful, supervised browser sessions. Not for high-volume pipelines; pair with proxies to go further.
The strongest unpaid anti-detect result we've measured, bought with real operational pain.
Run evidence
Every attempt below is backed by saved artifacts: request/response metadata, body content, and (for browsers) screenshots. Raw JSON and artifacts live in the runner data directory.
Run data not available in this build.