← Back to blog

Karpathy Uses Obsidian Web Clipper. Save Vault Does the Same Thing — With Claude Built In

· Save Team
aillmknowledge-baseclaudemcpkarpathysave-vaultwiki

On April 3, 2026, Andrej Karpathy posted what may be the most-shared AI workflow of the year. He calls it LLM Knowledge Bases: a folder of markdown files that an LLM compiles into a personal wiki, then queries on demand.

The post lit up X. VentureBeat, Analytics Vidhya, and DAIR.AI all wrote it up within 48 hours. Half of AI Twitter started building their own.

But if you read Karpathy’s own description carefully, the workflow has three pieces:

  1. A web clipper to convert pages into clean markdown (he uses Obsidian Web Clipper)
  2. A folder of .md files on disk
  3. An LLM with filesystem access that can read, link, and answer questions across the folder

That’s three tools to install, configure, and keep in sync. A surprising number of people have given up at step 3.

Save Vault collapses all three into a single Mac app and a single toggle.

What Save Vault Actually Is

Save Vault is the macOS companion to the Save Chrome extension. It runs as a menu bar app and does three things:

  • Receives markdown from the Save extension over localhost:7421 and writes it to ~/Documents/Save Vault/
  • Organizes files into knowledge bases (subfolders you create — General, AI Research, Competitors, whatever you need)
  • Exposes those files to Claude via a built-in MCP server, started with one toggle in the menu bar

That last bullet is the new part. And it’s the part that finally makes Karpathy’s pattern accessible to people who don’t want to edit JSON config files.

The “Connect to Claude” Toggle

Open Save Vault. Click the menu bar icon. Flip Connect to Claude on.

That’s it. The app writes the MCP server entry into your Claude Desktop and Claude Code config, restarts the server, and your entire vault becomes searchable from any Claude conversation.

Behind the scenes, Save Vault exposes four tools to Claude:

ToolWhat Claude can do
list_knowledge_basesSee all your KBs and how many files each contains
list_filesBrowse files inside a specific KB
read_fileOpen any saved markdown file
searchFull-text search across the entire vault

When Claude is connected, it always checks your Save Vault before answering. If you saved a Karpathy thread last week and ask “what did Karpathy say about LLM wikis?”, Claude reads your saved file and quotes it back to you — instead of paraphrasing from training data that might be outdated or generic.

Toggle it off and Claude goes back to being a generalist. No uninstall, no config editing, no broken state.

Side-by-Side With Karpathy’s Setup

Karpathy’s stack is powerful but assumes you’re comfortable with developer tooling. Here’s how Save Vault maps to each piece.

Karpathy’s stackSave Vault equivalent
Obsidian Web ClipperSave Chrome extension
raw/ folder of .md files~/Documents/Save Vault/{kb}/
Hotkey to download images locallyBuilt into the extension — images extracted with the page
Filesystem MCP server (manual JSON setup)One toggle in the menu bar
Multiple research projects in subfoldersKnowledge bases (each KB is a folder)
Wiki “compilation” pass with Claude CodeRun Claude Code in the vault folder, ask it to compile

The extension and the Mac app do steps 1–4. You bring Claude.

A Real Workflow, End to End

Here’s what the loop looks like once you have it set up.

Monday morning. You read three articles about agent-based code review, one Karpathy tweet thread, and a Hacker News discussion about MCP. You hit the Save button on each one. The Save extension generates clean markdown and ships it to Save Vault. They all land in your AI Research KB.

Monday afternoon. Open Claude Desktop. Ask: “What are the main critiques of agent-based code review from this week’s reading?”

Claude (with Save MCP connected) calls search on your vault, finds the three articles you saved that morning, reads them with read_file, and answers using your sources — with file names you can verify. No hallucination, no generic LLM blather.

Friday. You have ~30 files in AI Research. Open Claude Code in the vault folder and ask it to act as Karpathy’s “LLM wiki compiler”: read every file in the KB, write a _wiki.md index, generate concept pages, and link them with backlinks. Claude grinds through the folder and produces a navigable structure.

Next month. You have 150 files. The wiki has compounded into something no public LLM has access to: your curated, structured view of an entire research area. Every question you ask gets grounded in it.

This is exactly the loop Karpathy described. The only difference is that the Save Vault pieces are pre-wired together.

Why Bother When You Can Just Use Obsidian

A fair question. Obsidian + Web Clipper + a manual filesystem MCP server gets you most of the way there. People have been doing it for months.

A few reasons people switch to Save Vault:

  • The clipping is better. Save uses Gemini to extract clean content. It handles paywalled pages you have access to, YouTube transcripts, X threads, TikTok captions, and Instagram reels. Web Clipper struggles with anything that isn’t a static article.
  • No config files. Save Vault writes the MCP entry for you and tells Claude Desktop to reload. You never touch JSON.
  • No vault to manage. Save Vault is a folder, not a vault format. You can still open it in Obsidian if you want plugins and backlinks. But you don’t have to.
  • One toggle to disconnect. When you don’t want Claude grounding answers in your saved knowledge — say, you’re asking it to brainstorm freely — flip the toggle and you’re back to baseline.

If you already love Obsidian, keep Obsidian. Point it at ~/Documents/Save Vault/ and you get backlinks plus Save’s ingestion plus the Claude toggle. They compose.

The Bigger Idea

Karpathy’s post resonated because it pointed at something everyone using LLMs already feels: the model is smart, but it doesn’t know what you know. Every conversation starts from zero. Every insight you generate disappears when the chat ends.

A persistent markdown vault, ingested from the web with one click and exposed to Claude with one toggle, fixes that. Your AI starts with your context. Your reading compounds. Your research has a memory.

Save Vault is that — packaged so you can install it in five minutes instead of an afternoon.

Get Started

  1. Install the Save Chrome extension
  2. Install Save Vault from savemarkdown.co
  3. Open Save Vault → click Connect to Claude
  4. Save the next article you read
  5. Ask Claude about it

That’s the whole onboarding. The compounding starts with the first save.


Save Vault is free. The Save extension is free for 3 saves a month, $3.99/mo unlimited. Download both at savemarkdown.co.