cdsassj00/llm-wiki

llm-wiki is a skill that coding assistants like Claude, Cursor, or Codex can install and use to turn a folder of your documents into a browsable wiki with a 3D knowledge graph view, in a style inspired by Andrej Karpathy. You point an agent at a folder of documents such as lecture notes, research papers, or project files, and the agent asks a few questions, then builds an organized wiki out of them along with a visual map of how the documents connect. Installation happens through a command called npx skills add, which installs only the skill code itself to a shared agent skills folder. It does not read or copy your personal files, and it does not create any workspace or API keys on its own. The actual wiki setup and any AI configuration only happen afterward, inside a local workspace folder that you choose yourself. Once set up, a local bridge program runs only on your own computer, listening at the local address 127.0.0.1, and serves a handful of fixed actions: checking health, searching, asking questions, and exporting results. There is no general command execution endpoint. A one-click launcher script starts this bridge and opens an HTML page showing local search, document previews, links between documents, and an optional AI chat feature. By default, everything stays on your machine. If you turn on the optional AI search feature, only short excerpts from your top search results, capped at about 12 kilobytes, are sent to whichever AI provider you choose. Supported providers include the Claude command line tool, OpenRouter, OpenAI, Gemini, and Anthropic directly, and you only need to set up one of them. API keys are meant to be entered through a dedicated hidden-input setup tool rather than typed into chat windows or command lines. The project requires Python 3.10 or newer, plus a small set of Python dependencies, and optionally LibreOffice if you need to read older Office document formats. It is aimed at people already using AI coding agents who want a private, local way to organize and query their own documents, and it is released under the MIT license.