I built the "LLM Wiki" Karpathy sketched out, and open-sourced it
A few weeks back, Andrej Karpathy shared an idea in an April 2026 tweet and gist he called the "LLM Wiki": instead of chatting with your documents, an agent maintains a cross-linked wiki for you. It stuck with me, so I built it from scratch.
Here's the gap it fills. RAG chatbots (NotebookLM, ChatGPT file chat) are stateless. They re-read your files on every question and never leave you anything that accumulates. Note apps (Obsidian, Notion) are the opposite: you do all the writing, linking, and contradiction-checking by hand, so they never scale with how much you read.
LLM Wiki sits in between. You drop in a source (text, PDF, URL, image) and an agent does three things:
- Ingest: writes new pages and refactors existing ones, so each source enriches the whole wiki instead of just appending a page.
- Query: answers questions over the corpus with citations.
- Lint: finds contradictions, orphans, and stale claims, with one-click fixes.
The decisions I care most about:
- Everything is plain markdown on your own disk. Delete the app and your wiki still opens in Obsidian or vim. No lock-in.
- Local-first and bring-your-own-key. Run it completely free and private with local models via Ollama, or use frontier models through OpenRouter per operation.
- TypeScript end to end, ~194 tests, cross-platform, one-command npm install. I wanted it to feel finished, not like a demo.
Where it's at, honestly: MIT open source, live on npm, currently v1.2.3 with a 3D knowledge graph, multi-wiki support, and a first outside contributor who shipped the Ollama integration. Pre-revenue by design. There's a hosted version on a waitlist at llmwiki.cc that I'll turn into paid tiers if demand shows up, but the open-source local-first product stays free and intact.
A candid launch note for fellow makers: I went live across socials, Product Hunt, and Hacker News almost at once, and HN auto-flagged my Show HN within a minute. Lesson learned the hard way: pointing your existing social audience at a fresh HN post looks exactly like a voting ring to their detector. Stagger your channels and let HN rise on its own readers.
What I'd love from you:
- Try it: npm install -g @syasas/llm-wiki
- Tell me where it breaks. The part I'm least sure about: every ingest is a full refactor pass over the wiki, which is what makes knowledge compound, but I don't know how cleanly that holds up past a few hundred pages.
- Contributors welcome. There are labeled good-first-issues on the repo: github.com/ddsyasas/llm-wiki
Building this in the open, so any feedback is gold.
