About
A personal Wikipedia an LLM maintains for you.
You drop in sources — papers, articles, notes, URLs. An LLM agent reads them, writes cross-linked pages, keeps an index, and revises older pages when new sources change the picture. The result is a knowledge base that compounds: each source makes every page richer, not just one new page longer.
The pattern
Built on Karpathy's idea
In April 2026 Andrej Karpathy sketched a pattern for personal knowledge bases that treats the LLM as the knowledge-engineer rather than the search-engine. Three layers (raw sources, the LLM-maintained wiki, a schema the LLM reads on every call) and three operations (ingest, query, lint) form a loop that gets better as you feed it.
“Obsidian is the IDE. The LLM is the programmer. The wiki is the codebase.”
LLM Wiki is an open-source implementation of that pattern — local- first, bring-your-own-key, plain markdown all the way down. You can delete the app and your wiki keeps working.
What you get
The wiki + a 3D view of its shape
Your knowledge ends up in two complementary views:
- The wiki itself — markdown pages grouped by type (overviews, concepts, entities, comparisons, sources), with full backlinks and source-lineage trails. Read it like a textbook, search it, edit it.
- A 3D graph of every page and every cross-link — same engine and look as Obsidian's graph view, but colored by page type rather than free-form tag. Watch your knowledge literally grow over time as you ingest more sources; orphans and hubs become spatially obvious.
Why it exists
The gap nobody was filling
RAG chat (NotebookLM, ChatGPT files)
Stateless. Rediscovers your corpus from scratch on every query. Never accumulates anything you can read later.
Note-taking apps (Obsidian, Notion)
All the maintenance burden on the human. You write, you cross-link, you check for contradictions. Nothing scales.
LLM Wiki sits between them. The LLM does the maintenance; the wiki accumulates value; you own the files. After three months of feeding it, you have a navigable, cited, deliberately-organized body of knowledge about whatever you care about — without ever having written a page yourself.
One wiki, or several
A wiki per topic — switch with one click
A wiki is meant to focus on one topic — the schema you set on first run keeps the LLM on scope. For separate topics (say Physics, ML research, and a Personal knowledge base) you keep separate wiki folders and switch between them in Settings → Wikis. Switching re-points the whole app on the next request — no restart, no port juggling. Each wiki is its own folder you fully own.
Who it's for
If any of these describe you
- ResearchersSynthesize a literature you keep coming back to. Citations stay attached to claims, contradictions surface in lint.
- Lawyers / paralegalsBuild a working knowledge of a case or a regulatory area. Every claim traces back to the source document on disk.
- Doctors / cliniciansKeep a private reference of guidelines, studies, and your own notes. No cloud upload of patient-adjacent material.
- Journalists / analystsBeat-reporting where the same names, orgs, dates keep recurring. The wiki becomes your second brain for the story.
- EducatorsBuild a course outline that updates itself as you ingest new readings. Cross-references stay live.
- Indie hackers / technical foundersA private knowledge base for the market you're building in. Ingest competitor docs, customer notes, your own decisions.
Principles
What it refuses to do
Files over databases.
Every artifact you might want to keep is a markdown file in a folder you chose. SQLite is metadata only — safe to delete, regenerable.
No lock-in.
Open standards (Markdown, SQLite), bring-your-own-key for the LLM. Uninstall the app, your wiki still works in any editor.
The agent does the work.
You shouldn't have to manually update cross-references, generate summaries, or maintain an index. The LLM compiles; you read.
Honest costs.
Every operation tells you roughly what it will cost in tokens before it runs. No surprise bills.
Local-first, BYOK.
No telemetry, no remote storage, no auth. You bring your own OpenRouter key; we never see it.
Tech
The stack, briefly
- Framework
- Next.js 14 (App Router), TypeScript strict
- UI
- Tailwind + shadcn-style components, Fraunces / Crimson Pro / Inter
- Storage
- Plain markdown files + SQLite (better-sqlite3) for metadata
- LLM
- OpenRouter (BYOK) — Claude / GPT / Gemini / Llama, you pick per operation
- Search
- FTS5 on page bodies + frontmatter title/tags
- Ingestion
- mammoth (DOCX), officeparser (XLSX/PPTX), @mozilla/readability (HTML/URL), vision models for PDF/image
- License
- MIT
For the deep dive, see the Developers page or the /docs folder on GitHub.
Credits
Made by Yasas. Pattern by Karpathy.
Built by Yasas as a from-scratch implementation of the LLM Wiki pattern described by Andrej Karpathy. Released under the MIT license. Contributions and forks welcome on GitHub.
v1.2.3 · Install with npm install -g @syasas/llm-wiki. Verified on macOS, Linux, Windows.
Project website: llmwiki.cc — also has a hosted version (waitlist) for people who'd rather not install anything.
LLM Wiki is not affiliated with Andrej Karpathy or Anthropic; the pattern is his, the implementation is independent.