robiulhasan.de
● Live

Titan Search — Self-Hosted Private AI Search

A private, self-hosted search infrastructure that solves Google's CAPTCHA blocking of automated browsing — with zero tracking, unlimited queries, and an AI synthesis layer.

  • SearXNG
  • Python
  • Docker
  • AI
  • Datenschutz

Problem

Google blocks automated browsing with CAPTCHAs. For an AI Operating System that needs to research continuously, this was a hard wall:

  • Every automated research query eventually hit a CAPTCHA challenge
  • Browser automation against Google is unreliable and rate-limited
  • The system needed unlimited, private, reliable search — not a workaround

Solution

A fully self-hosted search stack:

  1. SearXNG (open-source metasearch engine) in Docker — aggregates results from multiple engines without tracking users
  2. Custom AI synthesis layer — wraps raw results with an AI-generated summary
  3. Private by design — no search history, no tracking, no third-party cookies

Architecture

AI Agents (Hermes, Claude, Gemini)
    │
    ▼
titan-search (synthesis wrapper)
    │
    ▼
SearXNG (Docker, 127.0.0.1:8888)
    │
    ▼
Multiple search engines (aggregated, untracked)

Key decisions:

  • SearXNG bound to 127.0.0.1 only — never exposed publicly (ADR-010)
  • Custom settings.yml branded “Titan Search”
  • UID permissions fixed for container security (not running as root)
  • Result: 42 results per query, no CAPTCHAs, unlimited volume

Live Evidence

Metric Result
CAPTCHA blocks 0 (vs constant with Google)
Results per query 42
Cost €0 (self-hosted, existing VPS)
Privacy No tracking, no history
Availability 24/7, no rate limits

Learnings

  1. Self-hosted beats workaround. Google’s CAPTCHA isn’t a technical challenge to solve — it’s a signal to own your infrastructure.
  2. Metasearch is the privacy answer. Aggregating engines means no single provider builds a profile of you.
  3. The AI layer changes everything. Raw search results are data; synthesized answers are intelligence.

Self-Critique

  • Search result quality varies by engine health — needs better failover
  • No caching layer yet — repeated queries re-fetch
  • Synthesis quality depends on the model used; could be improved with a dedicated ranking step

Evidence References

  • Repo: /root/2nd-brain (architecture docs)
  • Live: 127.0.0.1:8888 (private, Tailscale-only)