A web search engine. Kind of.
Every result is fabricated by a large language model, and every destination page is fabricated in full by a large language model, too. Nothing you find here is real, and that is the whole joke.
Made with ✨sparkles✨ by maragu: independent software consulting for cloud-native Go apps & AI engineering.
Contact me at markus@maragu.dk for consulting work, or perhaps an invoice to support this project?
Disclaimer: this project is 100% vibe coded.
- The homepage is a single search input.
- Submitting a query lands on
/?q=..., which shows ten Google-style result cards. Results are fabricated by Gemini 3 Flash Preview and cached in SQLite so the same query renders instantly on refresh. - Cards fill in progressively via Server-Sent Events driven by Datastar signals.
- Clicking a result blocks for up to two minutes while Gemini 3 Flash Preview fabricates a full standalone HTML document, which is then cached in SQLite and served raw.
Set your Google Gemini API key, then let the watch loop do the rest:
cp .env.example .env
echo "GOOGLE_API_KEY=AIza..." >> .env # replace with your key
make watchThe app writes logs to app.log, rebuilds on file changes, and listens on the address configured by SERVER_ADDRESS (default :8080).
Inspect the cache directly with sqlite3 app.db (remember pragma foreign_keys = 1;).