karpathy.app GitHub

Your notes. Your AI.
Everywhere you are.

karpathy.app puts your Markdown vault and an AI that reads, writes and ingests into it on your iPad, phone and Mac — synced through the GitHub repo you already have.

Built on Andrej Karpathy's LLM Wiki idea →

Get the code How it works

Your vault, your repo

A vault is a GitHub repo. Obsidian on your desktop and phone keeps working on the same files — no second sync system.

An AI that works in your notes

Chat with any model — Claude, OpenAI, a local one. It searches, reads and edits your Markdown, and you see every file it touches.

Your skills come along

Skills, AGENTS.md and CLAUDE.md in your vault work as they do in the terminal, now from the sofa.

You decide what gets committed

Your edits and the AI's stay uncommitted until you review them and commit. The AI never pushes on its own.

See it on every device

The demo vault on a Mac, an iPad and an iPhone: a wiki page next to a chat whose answer comes from the page the AI just read.

karpathy.app on a Mac: file tree, a wiki page about which coffee brewer suits which bean, and an AI chat that read the page and recommends the V60
Mac — files, page and chat side by side
karpathy.app on an iPad in landscape: the same wiki page and chat
iPad — the same, made for touch
karpathy.app on an iPhone: the chat tab with the question, the file the AI read and its answer
iPhone — the chat in your pocket

How it works

  1. Connect a GitHub repo as a vault.
  2. Edit and chat — raw Markdown with live preview, an AI at your side.
  3. Commit — one tap commits and pushes; Obsidian picks it up.

Where the idea comes from

karpathy.app is built on an idea Andrej Karpathy shared in April 2026: let an LLM turn your raw sources into an interlinked Markdown wiki, and keep it healthy.

LLM Knowledge Bases

“Something I'm finding very useful recently: using LLMs to build personal knowledge bases for various topics of research interest. …”

Andrej Karpathy on X, · View on X

The gist: LLM Wiki

Two days later, on , he wrote the pattern up as an “idea file”: raw sources go into a folder, the LLM ingests them into wiki pages, answers questions from the wiki and lints it.

Read the gist on GitHub

That loop runs in a terminal. karpathy.app brings it — and your wiki skills — to the iPad and the phone.

Code, not a service

karpathy.app is open source. We ship the code — not a running system. To use it, you run it yourself. That takes:

A server

A small Linux box (e.g. 2 vCPU, 4 GB RAM) that runs the docker compose stack and holds the clones of your vaults.

Tailscale

The app isn't on the open internet: your iPad, phone and Mac reach it over your own tailnet.

Lots of secrets, tokens and keys

A GitHub token for your vaults, an API key for your LLM provider, the app's access token, a DNS API key for the TLS certificate — and a safe place to keep them.

Rather have it hosted?

If you'd like karpathy.app as a hosted service, write me at till.gartner@gmail.com.

Write me View the source