LLM wiki on GitHub: Karpathy's gist and 12 repos compared

Updated October 6, 2026 · 9 min read

Karpathy’s LLM wiki is not a GitHub repository. It is a gist, llm-wiki.md, an “idea file” you paste into your own agent so it can build the wiki with you. Everything else on GitHub is built on it by other people: desktop apps, a VS Code extension, Obsidian plugins, agent skills and compilers. Below are 12 of them, checked with the GitHub API on 1 October 2026, and how to choose.

If you are new to the pattern itself (raw sources, a wiki the agent writes, a schema file, and the ingest, query and lint steps), start with our guide to Karpathy’s LLM wiki. This guide is about the code.

Where is Karpathy’s LLM wiki on GitHub?

At gist.github.com/karpathy/442a6bf555914893e9891c11519de94f. It has no code to install. The gist says it is “designed to be copy pasted to your own LLM Agent (e.g. OpenAI Codex, Claude Code, OpenCode / Pi, or etc.)” and that “your agent will build out the specifics in collaboration with you.”

That is why so many repositories exist. Each one is somebody’s answer to “the specifics”: which app holds the pages, which agent does the writing, and how the schema is packaged. None of them is official.

Which LLM wiki repos are on GitHub?

Stars, last push and license come from gh api repos/<owner>/<repo> on 1 October 2026. Descriptions come from each README, read the same day.

Note on nashsu/llm_wiki: the API reports no recognised license, but the repository’s LICENSE file and README say GNU GPL v3.0.

Two things stand out. Ten of the 12 were pushed to in September or later, so these are live projects that change often; check the README before you rely on any detail here. And stars do not track search rank: the plugin and the VS Code extension that Google ranks third and fourth for “llm wiki github” have 116 and 67 stars.

A GitHub search for “llm wiki” sorted by stars puts Tencent/WeKnora first, at 31,594. It is a broader knowledge platform (RAG, a reasoning agent and a self-maintaining wiki mode), so we left it out of the table.

Which LLM wiki apps work without a terminal?

nashsu/llm_wiki is the most-starred repository built on the pattern. It is a cross-platform desktop app with installers on its Releases page (latest v0.6.12). It ingests PDFs, Office files, EPUB, images and web clips, draws a knowledge graph with community detection, and can add vector search through LanceDB. It also runs a local HTTP API and MCP server on 127.0.0.1:19828, and ships a skill for Claude Code and Codex:

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

Pick it if you want one app that does ingest, search and the graph for you.

inkeep/open-knowledge (OpenKnowledge) is a WYSIWYG markdown editor for macOS, Windows and Linux, plus a web UI. It edits side by side with Claude, Codex, OpenCode and Pi, includes MCP and skills for LLM wikis, and shares with a team through git and GitHub. The README calls it “Private, local, and free.” Pick it if you want to read and edit the pages yourself in something nicer than a code editor.

npm install -g @inkeep/open-knowledge

microsoft/llmwiki is a VS Code extension. It adds Entities, Concepts, Raw Sources and Backlinks views to the sidebar and an @wiki participant in Copilot Chat. It needs a GitHub Copilot subscription, because it uses VS Code’s Language Model API. The README installs it by building a .vsix from source. Its @llmwiki/core package also runs an MCP server so Claude Desktop or Cursor can read and write the same wiki. Pick it if your team already lives in VS Code with Copilot.

Which ones work with Obsidian?

Karpathy keeps Obsidian open next to his agent, so Obsidian is the most common front end. Three approaches:

• GD4AI/obsidian-llm-wiki is a community plugin. In Obsidian, go to Settings, Community plugins, Browse, and search “Karpathy LLM Wiki”. No terminal and no Python. It supports 16 or more model providers, runs on desktop and mobile (Obsidian 1.11.4 or later), and retrieves through the link graph instead of embeddings.
• AgriciDaniel/claude-obsidian turns Claude Code (or another Agent Skills host) into the writer and keeps the vault as plain files. Setup is a git clone and a Python init command that shows a plan before it changes anything.
• kytmanov/obsidian-llm-wiki-local runs on Ollama by default, so nothing leaves your machine. You can reject a draft with a reason and the next compile takes it into account.

pip install obsidian-llm-wiki

Our guide to running an LLM wiki in Obsidian walks through the setup, and Obsidian with Claude Code covers the agent side.

Which ones are skills or plugins for coding agents?

These add the ingest, query and lint routine to the agent you already use. The wiki stays a folder of markdown in your project or home directory.

Astro-Han/karpathy-llm-wiki is the smallest: one skill in the Agent Skills format for Claude Code, Cursor, Codex and others. Sources go in raw/, pages in wiki/. Its examples/ folder comes from a wiki the author keeps daily (94 articles from 99 sources, per the README).

npx add-skill Astro-Han/karpathy-llm-wiki

praneybehl/llm-wiki-plugin targets the widest list of agents: Claude Code, Codex, Cursor, Gemini CLI, OpenCode, OpenClaw, Pi, OMP and Hermes. Search runs locally with no hosted vector database. Version 3.2.0 adds /wiki:learn and /wiki:evolve, and the README is candid that its own study showed no measured quality gain from the evolve step yet.

/plugin marketplace add praneybehl/llm-wiki-plugin
/plugin install llm-wiki@llm-wiki

For other agents: npx skills add praneybehl/llm-wiki-plugin -a <agent> -g.

SamurAIGPT/llm-wiki-agent follows the gist’s layout closely (index.md, log.md, overview.md, and folders for sources, entities, concepts and syntheses) and writes a graph/graph.html you can open in a browser. You clone it and tell your agent ingest raw/papers/some-paper.md.

nvk/llm-wiki goes furthest toward research: parallel multi-agent research, thesis-driven investigation, session memory, and a lint command with --fail-on exit codes for CI. It packages for Claude Code, Codex and OpenCode.

The schema file these skills write into is usually CLAUDE.md or AGENTS.md; our AGENTS.md guide covers what belongs in it.

Which ones compile a wiki in one pass?

atomicstrata/llm-wiki-compiler treats the wiki like a build output. You point the llmwiki CLI at papers, notes, transcripts, PDFs or web pages and it compiles typed pages with citations, review queues and freshness checks. It has a local viewer, an MCP server and an SDK, and exports to Open Knowledge Format, JSON-LD, GraphML and llms.txt.

npm install -g llm-wiki-compiler

Pick it when the sources arrive in batches and you want a repeatable pipeline rather than a conversation.

lucasastorian/llmwiki is a web app: a Chrome extension clips pages and highlights, and a nightly Claude Routine writes them into the wiki. Claude.ai, Claude Code and Codex connect over MCP. You can self-host it, and the README says you can try it free at llmwiki.app.

What does a finished LLM wiki look like?

Most of the repos above are tools, so their own files are documentation, not a wiki. To see the pattern filled in, look at a repository that is the wiki.

coleam00/cole-medin-knowledge-base is one: a wiki compiled from a YouTube creator’s long-form catalog, with concepts/, entities/, sources/, raw/ transcripts and an index.md. The README suggests dropping it next to a project so any coding agent can answer from it with citations to the video and timestamp.

You can open any of these as a graph in your browser with our free wiki graph viewer. Paste the repository, or use the address directly:

https://dexio.wiki/wiki-graph/?repo=github.com/<owner>/<repo>

For example, the knowledge base above as a graph. A link to one folder (github.com/owner/repo/tree/main/wiki) works too, which helps when the wiki sits inside a tool’s repository. Nothing is uploaded; it runs in the browser.

What we saw in 46 public LLM wikis

In September 2026 we looked at 46 public LLM wikis on GitHub: repositories with a folder of at least 20 markdown pages and at least 30 [[wikilinks]] between them. Together they held 7,195 pages, and the median wiki had 76.

Two things worth copying:

• Many are starter kits and demos. 13 of the 46 were templates or sat in example folders. If you want to see a wiki under real use, look for a recent log.md with many entries, not a high star count.
• Keep the wiki in its own folder. In 28 of the 46, the wiki sat at the repository root next to the project’s own README and docs, which mixes the two in search and in the graph. A wiki/ folder keeps them apart.

How do I pick one?

If you are unsure, the gist plus one skill is the cheapest start: everything is plain markdown, so you can move to any of the others later.

Share it with your other agents

Almost all of these keep the wiki on one machine, written by one agent at a time. That is the gist’s reference setup and it works. It gets harder when Claude Code on your laptop, Codex in CI and a chat app in the browser all need the same pages, because a folder has to be synced or merged between them.

Dexio is a hosted wiki that agents read and write over MCP. Claude Code, Codex, Cursor, Hermes Agent, OpenClaw, Claude and ChatGPT connect to one wiki of linked markdown pages, every change records which agent made it, and you see what your agents know as a page graph in the app. It is open source (AGPL-3.0), free for one person, and $10 (Team) or $20 (Business) a member a month for teams. Any wiki downloads as markdown, so you can still open it in Obsidian or any tool above.

To connect an agent that can run commands, tell it “Look at dexio.wiki and log me in.” Setup per agent is in our guides, for example Claude Code and Grok Bot.

If you want everything offline, or a single desktop app, one of the repositories above is the better pick.

FAQ

Is there an official Karpathy LLM wiki repository? No. Karpathy published the gist; the repositories are community projects. microsoft/llmwiki comes from Microsoft’s GitHub organisation but is also built on the gist.

Which LLM wiki repo has the most stars? Among repositories built on the pattern, nashsu/llm_wiki, with 20,132 on 1 October 2026.

Do I need Obsidian? No. The wiki is a folder of markdown files. Obsidian is a good reader for it, and so is VS Code, OpenKnowledge, or the wiki graph viewer.

Is an LLM wiki the same as RAG? No. RAG searches raw documents on every question; an LLM wiki writes the synthesis once and keeps it current. Some of the tools above add vector search on top. Our LLM wiki vs RAG guide compares the two.

Sources

• https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f
• https://github.com/nashsu/llm_wiki
• https://github.com/inkeep/open-knowledge
• https://github.com/microsoft/llmwiki
• https://github.com/GD4AI/obsidian-llm-wiki
• https://github.com/AgriciDaniel/claude-obsidian
• https://github.com/kytmanov/obsidian-llm-wiki-local
• https://github.com/Astro-Han/karpathy-llm-wiki
• https://github.com/praneybehl/llm-wiki-plugin
• https://github.com/SamurAIGPT/llm-wiki-agent
• https://github.com/nvk/llm-wiki
• https://github.com/atomicstrata/llm-wiki-compiler
• https://github.com/lucasastorian/llmwiki
• https://github.com/coleam00/cole-medin-knowledge-base
• https://github.com/Tencent/WeKnora
• https://dexio.wiki/wiki-graph/
• https://dexio.wiki/llms.txt
• https://dexio.wiki/agents.md