LLM Wiki

Implements Andrej Karpathy’s llm-wiki.md pattern (nashsu/llm_wiki). Not an Anthropic product.

What it does

Traditional RAG retrieves the source and regenerates an answer on every question, re-deriving the same knowledge from scratch each time. Andrej Karpathy’s llm-wiki.md pattern flips that: the LLM reads your documents, compiles them into a wiki once, and incrementally keeps it current.

LLM Wiki is a full cross-platform desktop application built on that pattern. Drop in documents and the LLM generates entity and concept pages with cross-references into a persistent knowledge base, and a built-in local HTTP API plus an agent skill let external agents like Claude Code query that wiki.

Features

• Two-step chain-of-thought ingest

  The original pattern read and wrote in a single step. LLM Wiki splits it into analysis (extract entities, concepts, and contradictions with existing knowledge) and generation (write wiki pages) across two LLM calls for better quality. A SHA256 cache skips unchanged files to save tokens.

• 4-signal knowledge graph with community detection

  Relevance is computed from direct links, source overlap, Adamic-Adar, and type affinity, and Louvain community detection discovers knowledge clusters automatically. Graph Insights surface surprising connections and knowledge gaps with one-click Deep Research.

• Vector semantic search and multimodal ingest

  Optional embedding-based retrieval via LanceDB (any OpenAI-compatible endpoint), plus extraction of embedded images from PDFs with vision-LLM captions surfaced in image-aware search.

• Deep Research and Chrome Web Clipper

  Multi-query web search via Tavily, SerpApi, or SearXNG auto-ingests results into the wiki, and a Chrome extension captures web pages with one click.

• Obsidian compatibility

  The wiki directory works as an Obsidian vault, using [[wikilink]] syntax, YAML frontmatter, index.md, and log.md.

Claude Code integration

LLM Wiki ships a built-in local HTTP API (http://127.0.0.1:19828, token-protected, 127.0.0.1-only) so external agents like Claude Code can query your wiki. A companion agent skill lives in its own repo and installs with one command:

npx skills add https://github.com/nashsu/llm_wiki_skill.git --skill llm_wiki_skill

After install, the agent can answer prompts like “what does my LLM Wiki say about X”, “show the neighborhood of node Z in my wiki graph”, and “rescan my wiki sources” by talking to your locally-running app — read-only by default, citing wiki page paths so you can verify in-app.

The skill has trigger discipline: it intentionally does not fire on generic “search my notes” or “check my Obsidian / Notion / Logseq” — only when you explicitly name LLM Wiki, my wiki, or 知识库.

Quick start

1. Download the desktop app from Releases (macOS .dmg, Windows .msi, Linux .deb/.AppImage).
2. Launch the app, create a project (choose a template), and configure your LLM provider (API key + model) in Settings.
3. Import documents (PDF, DOCX, MD, etc.) under Sources and watch the Activity Panel build wiki pages.
4. Enable the local API and generate a token in Settings → API Server.
5. Install the companion skill into Claude Code and query your wiki in natural language.

Notes

• Not an Anthropic product — a community implementation of Karpathy’s llm-wiki.md pattern (nashsu/llm_wiki).
• GPL-3.0 license — free to use, modify, and redistribute.
• App and Claude Code integration are separate — the integration goes through a skill in a separate repo (nashsu/llm_wiki_skill), which does not declare a license.
• LLM provider is separate — choose among OpenAI, Anthropic, Google, Ollama, or Custom and supply your own API key. Anthropic is one of the supported providers.
• Local API is localhost-only — it runs on 127.0.0.1:19828 and is token-protected, so the integration works only between the app and an agent on the same machine.