No. The vault is plain Markdown, and your agent reads and writes it directly. Obsidian is the nicest way to browse it yourself: backlinks, graph view, and a dashboard that needs the free Dataview plugin.

Which agents does it work with?

Claude Code and Codex, both tested end to end. Gemini CLI reads the same skills folder as Codex but is not tested yet. Any other agent that reads an instructions file can use it through a short block the installer adds with your permission.

Mac, Windows or Linux?

Windows and macOS. Automated installer and extractor tests run on both with every change. Linux is not tested.

What does it cost to run?

The kit is free. Captioned YouTube videos, web pages and PDFs cost nothing to add. Audio files and videos without captions go through OpenAI's Whisper with your own key, about $0.006 a minute of audio.

Will it overwrite my existing wiki?

No. If you built a wiki from Karpathy's gist, the installer finds it, asks you, and links it. Your pages stay exactly as they are and your agent searches them next to the vault. The installer checks your files before and after, and stops if anything changed.

Does my data leave my machine?

Your vault stays on your machine. Your agent sends what it reads to its model provider, as in any agent session. Sources are fetched from where they live, and audio goes to OpenAI only when you add audio or a video without captions. Contact files are confidential by default, and journal entries become confidential when they mention a name on your private list.

Can I use it at work and change it?

Yes. It is MIT licensed: use it, change it and share it, as long as the copyright notice stays in. The packages it installs keep their own licenses. PyMuPDF, which reads PDFs, is AGPL-3.0, so check its terms before you build it into something you distribute.

When a new version ships, you get the new zip by email. Give it to your agent with the same install sentence. Skills you never changed are updated, skills you changed are kept with the new version waiting beside them, and your notes are never touched.

Knowledge Vault Kit: Karpathy's LLM Wiki, built out with video, journal and CRM

The Knowledge Vault Kit is a free, working version of Andrej Karpathy's LLM Wiki that your coding agent installs for you. Save a YouTube video, an article, a PDF or a voice memo, and it becomes a cited, cross-linked page in a folder of plain Markdown files on your own machine. Nine skills do the filing and answer your questions from what you saved. Three of them go past the gist: a video pass that reads what is on screen, a journal that answers you from your own sources, and a personal CRM whose people stay confidential.

Karpathy's LLM Wiki, and what this kit adds

In April 2026, Karpathy published LLM Wiki, an idea file you paste into your own agent. It describes three layers (your raw sources, a wiki the model writes, and a schema file with the rules), three operations (ingest, query and lint), and an index file and a log file. The sources it has in mind are documents such as articles and papers, and Obsidian is where you browse the result. It is intentionally abstract. It describes the idea and ships no code, and your agent builds the specifics with you.

This kit builds it out. What it adds:

Working skills, and an installer your agent runs for you. You answer its questions.
Video. YouTube transcripts, Whisper when a video has no captions, and an opt-in visual pass that reads key frames, copies on-screen prompts and code word for word, and links each timestamp to that moment in the video.
More source types. Web pages, PDF files, audio files, local Markdown and text files, and batches of YouTube videos.
A controlled topic list. New topics arrive as proposals, and nothing joins the list until you approve it.
Topic pages rebuilt in full from every source tagged with them, with disagreements between sources surfaced instead of smoothed over.
Saved answers that cite every claim.
A journal for reflections and voice memos. An entry that mentions a name from your private list is marked confidential, and your agent answers with advice drawn from your own sources.
Contacts, a personal CRM. One file per person, confidential by default, and every name goes on the journal's private list.
Voice notes, transcribed and cleaned up.
A health check that edits none of your pages and writes a dated report.
Upgrades that never touch your notes, and keep-and-link for a wiki you already have.

Built on his pattern, not affiliated with or endorsed by him.

How it works

You talk to your agent. The skills do the bookkeeping.

Add a source. Paste a link or a file path and say "add source". Your agent files it in a bucket, a category you name such as ai-coding-agents or roman-history. It checks for duplicates before and after extracting, then writes one page for the source and keeps the full text beside it.
Enrich it. The agent reads the full text and writes 5 to 10 key extracts. It picks 3 to 6 comparison fields for each new bucket, because a founder interview and a cooking video get compared on different things. Topics come from your list: one that isn't on it becomes a proposal for you to approve. Names the agent isn't sure of get flagged for you to check.
Rebuild the topic pages. Each topic page is rewritten from scratch from every enriched source tagged with it. You get a summary, the patterns with their evidence, and a Tensions section wherever sources disagree. A page built on three sources or fewer says its patterns may not hold.
Ask. Say "what does the vault say about X?" The agent reads the topic pages first, then the source pages. The skill tells it to cite a source for every claim and to drop anything it cannot cite. A substantial answer can be saved as a page, and next time the agent checks your saved answers first.

Most days you only add sources. Enriching and rebuilding topics is a weekly sit-down.

About once a month, say "lint the vault". It looks for untagged sources, contradictions a topic page missed, claims newer sources have overtaken, and missing cross-references. What to fix goes into a dated report, with the command to run for each.

Your own material lives in the same vault:

Journal. Start a message with "Journal:" or point your agent at a voice memo. The entry is saved with its date, and when your topic pages and sources speak to it, the agent answers with what they say.
Contacts. Start a message with "CRM:" and a name. The agent keeps one file per person, confidential by default, and adds the name to the journal's private list, so journal entries that mention that person are marked confidential.
Voice notes. Point your agent at a recording. It transcribes it with Whisper, cleans it up into paragraphs, and pulls out key ideas and action items.

Journal entries and contacts never reach a topic page. Topic pages are built from your sources only.

What's inside the kit

Nine skills: add-source, enrich-source, promote-topic, rebuild-topics, query-vault, lint-vault, journal, crm and voice-note.
The installer, install.py, which your agent runs twice: first as a dry run that changes nothing, then for real after your yes.
INSTALL.md, the step-by-step brief your agent follows.
HANDBOOK.md: every command, what each source type costs, how fields and topics work, troubleshooting and updates.
An empty vault to start from, with Obsidian settings, a Start Here page and a Dashboard of five live views.
A README, a changelog and the MIT license, plus a Python requirements file and an example keys file where every key is optional.

Install

Your agent does the install. You unzip one folder, say one sentence, then approve a plan and your starter topics.

Unzip the folder into your workspace, the folder you open your agent in.
Tell your agent: "Install the Knowledge Vault Kit by following knowledge-vault-kit/INSTALL.md".
Approve the plan. Your agent checks for Python 3.11 or newer and offers to install it if it is missing. Then it shows what the kit will add and how many of your existing files change, normally none. Nothing changes until you say yes.
Approve your starter topics. Answer two short questions: what you plan to save, and a question you would want it to answer in a few months. The agent drafts 10 to 15 topics, says which answer each came from, and changes the list until you approve it.

Along the way, the agent offers to install ffmpeg, which audio and the visual pass need, and tells you which line of the keys file to fill in for any optional key. It never asks for a key in the chat. When it finishes, start a new agent session, paste your first link and say "add source".

Already have a wiki, or an earlier version

A wiki you built from Karpathy's gist. The installer looks for what his gist describes: an index.md with a log.md, or a raw folder beside a wiki folder. When it finds one, it stops and asks before changing anything. Say "keep and link", and nothing of yours is edited, moved or renamed. The vault goes in a new folder inside your wiki. After that:

Your agent searches your wiki pages after the vault's own, and cites them by file path.
The health check leaves them alone.
Your starter topics are drafted from your page titles.
Your raw Markdown and text files can come in as sources, a few at a time, with the originals untouched.

An earlier version. Give your agent the new zip with the same sentence.

A skill you never changed is replaced with the new version.
A skill you edited, or one from an earlier bundle, stays exactly as it is. The new version waits in the vault's .kit/pending folder for you to compare.
Your sources, topics, journal, contacts and saved answers are never touched.
Running the same version twice changes nothing.

We ran this upgrade on an earlier bundle. It added the vault files that were missing and changed none of the files already there.

Agents and systems

Windows: tested end to end in Claude Code and in Codex. Automated installer and extractor tests also run on Python 3.11 and 3.12 with every change.
macOS: covered by the same automated installer and extractor tests, on Python 3.12. YouTube downloads and the visual pass on a real video are not tested on a Mac.
Gemini CLI: reads the same skills folder as Codex. Not tested.
Linux: not tested.

Running costs and honest limits

Python 3.11 or newer. Your agent offers to install it.
An OpenAI key, only for audio files and videos without captions: about $0.006 a minute of audio, through Whisper. Captioned YouTube videos, web pages and PDFs cost nothing.
ffmpeg for audio files, videos without captions and the visual pass.
YouTube sometimes blocks caption requests. An optional proxy setting in the keys file works around it.
The Dashboard needs Obsidian's free Dataview plugin. Nothing else needs Obsidian.
The visual pass needs an agent that can view images. If yours can't, the pass stops and the source page stays as it was.
It is not impressive on day one. One source is worth about as much as a bookmark. The vault gets useful at around 10 to 15 sources per topic.
Your agent sends what it reads to its model provider, as with any file you open in an agent session. The kit runs no server of its own.
A few inputs need a step first. Audio over 25 MB has to be split, and a scanned PDF needs text recognition (OCR). Google Docs and pasted text are not supported yet.

Who it's for, and who it isn't

For people who already work in a coding agent and save more videos, articles and PDFs than they will ever remember, and for anyone who built a wiki from Karpathy's gist and wants working skills around it without starting over.

Not for anyone without a coding agent, since the agent does the install and all the filing. If you want a phone app or a hosted service, this isn't it: the vault is a folder on your machine. And three saved videos won't give you deep answers. The vault pays off as it grows.

Updates, license and who made it

Lifetime updates: when a new version ships, you get the new zip by email, and your agent upgrades in place with the same sentence you installed with.

MIT licensed: use it at work, change it and share it, as long as the copyright notice stays in.

Made by NocodePro, a startup studio that has run this vault since April 2026.

Running more than one project with an agent? The Agent Workspace Starter Kit sets up the workspace this vault lives in.

Knowledge Vault Kit