Mullvad: Shutting down our search proxy Leta

Original Article ↗ Hacker News Discussion ↗

What Leta Was and Why It’s Gone

  • Leta acted as a privacy proxy in front of Google/Brave (and possibly others), stripping tracking while returning their search results.
  • Several commenters always felt it was “on thin ice” because it apparently used Google’s API and cached results for ~30 days, likely conflicting with Google’s terms that restrict caching.
  • Some speculate Google’s tightening around automated access made the service non‑viable; others see Mullvad’s shutdown as a pragmatic decision to focus resources where privacy work has more impact.

VPN/Browser vs. Search Proxy

  • Mullvad suggests similar privacy can be achieved with a VPN plus a privacy‑focused browser; some find this reasonable, others argue that’s not a real replacement for a search proxy.
  • Debate over whether Mullvad could simply “scrape Google via VPN” ends with concerns about IP blocking and legal/ToS risk.

SearXNG and Self‑Hosting

  • Leta was popular as a backend for self‑hosted SearXNG. Its removal disappoints users.
  • Public SearXNG instances are widely reported as unreliable: rate‑limited, error‑prone, or returning irrelevant/foreign‑language results.
  • Self‑hosting SearXNG (often via Docker + Redis/Valkey) is described as relatively easy and more reliable, though individual providers still drop out occasionally.

Perceived Decline of Search Quality

  • Multiple people report DuckDuckGo becoming intermittently “unusable,” failing even on basic queries, or drowning in SEO/AI junk. Others say it’s improved and now offers per‑site blocking.
  • Many feel all search engines have degraded: more spam, AI‑generated slop, and non‑indexed niche content. Some suggest the web itself has hollowed out (forums gone, content paywalled/centralized), so there’s “less worth indexing” at all.

LLM‑Based Search: Usefulness vs. Risks

  • One camp claims users are shifting heavily to LLM‑style answers; another vehemently disagrees and cites screenshots where AI overviews confidently affirm mutually contradictory claims (e.g., “NFL viewership up” and “down”).
  • Concerns include:
    • “AI sycophancy” reinforcing user biases.
    • Non‑deterministic, hard‑to‑verify answers presented with undue confidence.
    • Safety risks (e.g., people doing hardware repairs based solely on AI instructions).
    • Potential to “kill websites” by diverting traffic to summaries, undermining incentives to publish new content.
  • Others counter that snippets and summaries have always reduced clicks, that LLM summaries are genuinely useful for many tasks, and that responsibility for verification still lies with users.

Alternatives and Trade‑offs

  • Kagi is widely praised for high‑quality, mostly Google‑sourced results, but criticized for using Yandex: some worry about indirectly funding Russia and about queries hitting Russian infrastructure.
  • Brave Search gets positive reviews; a Brave employee emphasizes it now runs a fully independent index. Users like the option to disable AI summaries.
  • Ecosia, Yandex via Tor, and others are also mentioned, each with privacy or geopolitical caveats.
  • Several people conclude that if we want non‑enshittified search, we likely have to pay for it.