WizSearch turns any body of knowledge into a shareable search URL. Drop that URL into Claude, GPT, Gemini — any AI — and it becomes an instant, queryable knowledge base. Built entirely by talking to Claude, over a two-day weekend.
I'm Eugene Gordeev, founder of Labs — a venture studio building AI-native companies from first principles. You can find me on X as @egordeev. I build through conversation — chat is my entire stack. This project came together with exactly two tools:
Pure chat and a repository. That's it.
I have retinitis pigmentosa, type RP25 — a genetic condition that narrows the visual field. Working through voice and chat instead of IDEs, dashboards, and terminals isn't a workaround for me; it's the native mode. That a person with failing vision can build something this involved, by talking, is a real blessing of this moment in AI.
Everything on this page — the engine, three demos, the crawler, the 2,000+ item index — was built inside a two-day weekend sprint. Roughly 24 hours of actual work.
Runs fully on your machine. Your index, your corpus, your queries — none of it leaves the browser. Works with sensitive data that should never touch a cloud.
Everything happens in the browser. The index is static JSON, the search is plain JavaScript. Entirely client-side, from first query to last.
GitHub Pages hosts it for free, forever. Free to host, free to query, free to scale.
GitHub Pages, S3, Cloudflare Pages, even IPFS. If it serves files, it serves WizSearch. Upload the folder and you're live.
If you can upload a zip to GitHub, you can run this. A GitHub account is the entire prerequisite; the whole install is drag-and-drop.
The crawler, the index, the engine, and this page — all written by talking to Claude in a browser tab. Chat is the entire build environment.
Text, PDF, Markdown, HTML, code, scraped web pages, transcripts. If the crawler can read it, the index can hold it.
Metadata extraction, structure inference, content cleanup — all during the crawl. The index holds more than the source files.
One interface serves all three. People browse, agents query programmatically, LLMs ingest the JSON directly.
Claude, GPT, Gemini, Llama, Mistral, your own fine-tune — all consume the same index. Switch models whenever you like; the index stays put.
Paste the search URL into any AI chat — Claude, GPT, Gemini — and the model has a live knowledge base, instantly.
Use it as a retrieval layer to augment any LLM with your own corpus. Point, query, inject — the URL itself is the retrieval layer.
WizSearch turns a body of text into a fast, semantic search engine that lives on a static web page. The whole architecture is a JSON index and a little JavaScript, running in the browser. Each demo points the same engine at a different corpus.
Every demo accepts a ?q= parameter. Append a query to any demo URL and it loads with that search already run. Example: demo3/?q=github opens the page and immediately shows matches for github. That makes results shareable — drop a link, the other person sees what you saw.
The design has a second, more interesting property: drop any of these URLs into any AI — Claude, GPT, Gemini — and it will work with the collected index. The page is static, the index is public JSON, and the search is deterministic. An AI that fetches the page gets a knowledge base on the spot.
The foundation. Indexes Anthropic's own Agent Skills documentation — the canonical, first-party source. This is the corpus that proved the pattern worked, and every later demo is built on what this one taught.
Same engine, different corpus. Forty-nine years of Buffett's annual letters indexed as one searchable body. Ask about moats, float, capital allocation, mistakes — you get the exact passages, across every year.
The furthest-along version of WizSearch. Indexes the full Claude ecosystem — every official Anthropic skill, MCP reference servers, and the 2,000+ community servers from awesome-mcp-servers — in a single shardable corpus. Built so Claude itself can validate: given a task, search here and decide which skill or MCP fits. Grows with the community.
Architectural features coming in v3:
?q=; any result view is a permanent linkOriginal engine archive. Smallest, cleanest version.
Current engine archive. Use this one by default.
The crawler. You start with a short onboarding conversation in any LLM — Claude, Grok, Gemini, ChatGPT — and the crawler produces a curated list of sources, file formats, and scraping strategies. That list then feeds the index-building process.