lucasastorian/llmwiki

LLM Wiki – an AI‑maintained personal knowledge base

What it is – LLM Wiki is a self‑hosting tool that lets you turn the files you already read (PDFs, Word docs, slides, notes, web clippings, etc.) into a searchable, cross‑linked “personal Wikipedia”. The heavy lifting is done by Anthropic’s Claude model, which you connect to via the MCP protocol. A nightly Claude routine reads any new material you’ve added, extracts the text, and updates or creates markdown pages, keeping the wiki up‑to‑date without you having to edit it manually.

Key components

• CLI (./llmwiki …) that indexes a folder of your existing documents, starts a local API server and a Next.js web UI, and generates a wiki/ folder of markdown pages.
• Chrome extension that lets you clip web pages or PDFs, highlight sections, and add comments; the saved data lands in the same workspace so Claude can ingest it.
• MCP server that exposes the workspace to Claude (Claude Desktop, Claude Code, or Claude Cowork). Claude can then use a small set of tools (search, read, create, edit, append, delete, lint, …) to read sources and write markdown pages.
• Backend abstraction – locally the data lives on your filesystem with a SQLite index; the same code can run in hosted mode using Postgres + S3, so the Claude tools behave identically.
• Graph viewer & visualizations – the web UI shows how concepts link together and can render Mermaid diagrams, charts, etc.

How you use it

1. Install – Python 3.11+, Node 20+, optional LibreOffice for Office files. Clone the repo, create a virtual‑env, install Python deps (api/requirements.txt, mcp/requirements.txt) and npm deps in web/.
2. Point it at a folder containing all your research files. The CLI creates a hidden .llmwiki/ index and a wiki/ folder for generated pages, but never moves or uploads your original files.
3. Start the service (./llmwiki open <folder>). The API, MCP server, and web UI become available at http://localhost:3000.
4. Connect Claude – run ./llmwiki mcp-config <folder> and paste the printed JSON into Claude’s config file. Then tell Claude to “ingest my sources and start building the wiki”.
5. Schedule a routine – create a Claude Routine (or a local scheduled task) that runs nightly with a prompt like:

   “Read the guide. Find everything added since the last run and update the wiki, creating new pages or folding material into existing ones, fixing citations and cross‑references.” The routine automatically keeps the wiki fresh.

6. Add content – either drag‑and‑drop files into the web UI, drop them directly into the workspace folder, or use the Chrome extension to clip and annotate web content.

Supported source types – PDFs (with optional Mistral OCR), Word/PowerPoint (via LibreOffice conversion), Excel sheets, HTML pages (converted to markdown), plain‑text files (.md, .txt, .csv, .json, …), and images (stored inline, readable by Claude when asked).

What you get

• A set of markdown pages under wiki/ that you can edit manually, version‑control, or let Claude modify.
• A searchable index (.llmwiki/index.db) that powers full‑text search and citation‑graph queries.
• Cross‑linking between pages and back to the original sources, plus visual graph and diagram support.

Roadmap highlights – adding ingestion from Slack, meeting‑note services (Granola), email forwarding, a public ingest API with webhooks, and Zapier integration to capture content from the tools you already use.

License – Apache 2.0.

---

Bottom line – LLM Wiki is a concrete, open‑source implementation of the “AI‑maintained personal wiki” idea popularized by Andrej Karpathy. It combines a simple local workflow (folder‑based storage + SQLite) with Claude‑driven autonomous updating, a Chrome clipping extension, and a modern Next.js UI, making it a practical way to build a second brain that grows automatically as you read and research.