Guide · LLM wiki without code
Karpathy's LLM wiki without code
You can run Karpathy's LLM wiki pattern without writing code or running an agent in a terminal: save articles and PDFs to Pith, and its wiki agent writes and maintains markdown pages with numbered citations, which you can then query from Claude or ChatGPT via MCP. This guide walks through it step by step, and shows the do-it-yourself route for comparison.
By Stefan Hartmann · Updated
What does the original LLM wiki set-up need?
Andrej Karpathy’s “LLM Wiki” gist from April 2026 is written for a coding agent. To run it as described you need:
- A coding agent such as Claude Code or OpenAI Codex, running on your machine with access to a folder.
- A folder layout with your raw sources and the wiki the agent writes, plus index.md and log.md.
- A schema file (CLAUDE.md or AGENTS.md) that tells the agent how to ingest, answer and maintain.
- A way to get sources in, for example the Obsidian Web Clipper to turn web pages into markdown.
- Optionally Obsidian as a viewer, git for history, and a local search tool once the wiki grows.
None of this is hard for a developer. For an analyst, researcher or consultant who just wants a cited wiki of what they read, the terminal, the schema file and the plumbing are the obstacle. The rest of this guide gets you the same loop — ingest, write pages, ask, keep current — in a browser.
How do you build an LLM wiki without code? Step by step
Step 1: Create a Pith account
Sign up with an e-mail address; no credit card is needed. Pith has no subscription — you pay per AI action from a prepaid credit, and new accounts get a one-time start credit with their first save plus a free credit every month. Details are on the pricing page.
Step 2: Check that saves use AI
A source only flows into the wiki when it is saved with AI. On Settings → Capture, the “Save with AI” switch controls this for every way of saving — the app, the browser extension, your phone, forwarded e-mails and your AI assistant. Saving with AI costs 2 ct per source; it adds a summary, facts, quotes, entities and tags, and hands the source to the wiki agent.
Step 3: Save your first sources
Paste a link that starts with https:// into “Save or search…” at the top of the sidebar (⌘K / Ctrl+K opens it), or use the field on the Saved page. PDFs work too; their pages are recognised at 1 ct per page. Other ways in, all listed on Settings → Capture:
- the browser extension for Chrome, Edge and Brave (installed as described on the Capture page; it is not in the Chrome Web Store yet),
- an iPhone or iPad Shortcut that adds “Save to Pith” to the share sheet, and the share menu on Android,
- forwarding a newsletter to your private Pith address,
- importing a bookmark export (HTML or CSV) from your browser or another read-later tool on Settings → Import.
Start with ten or so sources on one topic you actually follow. Karpathy’s advice applies here too: a narrow start with sources you trust beats a bulk import.
Step 4: Open the wiki and read what was written
Open Wiki. For each saved source, the wiki agent has decided which concepts it touches: it created pages for new concepts, patched existing pages where the source adds something, and attached the source as a further citation where it adds nothing new. Each page has numbered references in the text and a Sources list at the bottom, linking to the original and to your saved copy — if the original disappears, the saved text stays. A cross-linking pass adds related pages. Each page Pith writes or updates costs 5 ct; reading is free.
Step 5: Review and steer
This is the step Karpathy does by staying involved during ingest. In Pith you do it afterwards:
- Open a few of the sources behind a page and check that they say what the page says.
- A chip like “2 source conflicts” means two sources disagree; it opens the provenance trail with both passages side by side.
- Dismiss a page you don’t want — the agent will not recreate it.
- A source that was left out? Use “Add to wiki” on the bookmark, or in the wiki’s “No page yet” list. Code repositories, video and social sites and tool pages are kept out by default; the domain list is under Settings → Workspace → Wiki rules.
Step 6: Connect Claude or ChatGPT via MCP
In Claude or ChatGPT, add a custom connector with the server address https://mcp.pithlab.app/mcp, then click “Allow access” on Pith’s consent page. Access applies to one workspace. Other clients, such as Cursor, use a token from Settings → API tokens (“Generate MCP token”; it is shown only once). Reads and wiki writes over MCP are free. More on the MCP feature page.
Step 7: Ask questions from your assistant
Now you have Karpathy’s query step without a terminal. Prompts that work well:
- “Search my Pith wiki for [topic] and answer with the sources.”
- “Which of my saved sources disagree about [question]? Quote both.”
- “Has the source behind this claim been edited since I saved it?” (Pith can tell, because it kept the version you read.)
- “Write a wiki page comparing [A] and [B] from my saved sources.”
The last one is Karpathy’s “file good answers back” move. Your assistant can write a page to your wiki; Pith checks before storing it that each citation points to a source saved in your workspace, so an assistant cannot cite something you never saved. You can also have it list a page’s versions and restore an earlier one.
Step 8: Keep it growing
Keep saving as you read. Add RSS feeds on the Feeds page; their entries arrive in the Inbox and the morning report, where “Into the wiki” sends a source to the wiki. A daily health check flags pages without sources, orphan pages, pages not updated for 90 days and near-duplicates — Karpathy’s lint step, run for you.
What does a no-code LLM wiki cost?
With Pith you pay for what the AI does, from a prepaid credit, with no subscription and no seats:
- saving a source with AI: 2 ct,
- a wiki page written or updated by Pith: 5 ct,
- a PDF or image page recognised: 1 ct,
- a chat question to your wiki inside Pith: 2 ct.
Reading and searching the wiki, sharing and exporting are free, and so is everything your own AI assistant reads or writes over MCP. Failed jobs cost nothing. See the pricing page for the full, current list.
How do you build Karpathy's LLM wiki yourself instead?
If you prefer files you own and a workflow you design, the DIY route is a good one. A minimal start:
- Create a folder with raw/ for sources and wiki/ for pages, and run git init.
- Paste Karpathy’s gist into your coding agent and ask it to propose a schema file (CLAUDE.md or AGENTS.md) for your domain: page types, naming, how to cite raw files, what to do on ingest.
- Clip your first article into raw/ (the Obsidian Web Clipper converts web pages to markdown) and ask the agent to ingest it: summary page, affected concept pages, index.md, log.md.
- Read the result in any editor or in Obsidian, and correct the schema where the agent went wrong.
- Ingest one source at a time for the first few dozen, then batch.
- Once a week, ask the agent to lint: contradictions, stale claims, orphans, missing pages.
The things to plan for early are the ones that bite later: citations that point to raw files and survive rewrites, a rule to patch rather than rewrite, commits after every ingest, and a note on the page when two sources disagree. The LLM wiki guide explains each in more detail.
Pith or DIY: which should you pick?
Pick DIY if you want plain markdown in your own repository, enjoy shaping the schema with your agent, want to use a local model, or work mainly from chat logs and meeting notes. Pick Pith if you want capture from browser, phone and e-mail, citations that are checked against saved sources, conflicts and versions handled for you, and a wiki you can open in the browser and share as a link.
They also combine: a DIY agent can read Pith over MCP as one of its sources, so the wiki in your repository can cite what Pith has already captured and checked. If you are still deciding whether you need a wiki at all, read LLM wiki vs RAG first, or look at a real Pith wiki in the demo.
Frequently asked questions
Can I use Karpathy's LLM wiki without coding?
Yes. The gist is written for a coding agent, but the pattern — sources in, maintained and cited pages out, questions against the pages — does not require code. Pith runs it as a hosted service: you save links and PDFs, the wiki agent writes the pages, and you ask from the browser or from Claude or ChatGPT via MCP.
Do I need Claude Code or Codex for an LLM wiki?
Only for the DIY route. With Pith, the wiki agent runs on Pith's side; Claude or ChatGPT are optional and connect over MCP to read and write the wiki.
How do I connect Pith to Claude?
In Claude, add a custom connector with the server address https://mcp.pithlab.app/mcp and click “Allow access” on Pith's consent page. ChatGPT works the same way; clients such as Cursor use an MCP token from Settings → API tokens.
Can Claude write pages into my Pith wiki?
Yes. Over MCP your assistant can write wiki pages for free. Pith stores them only if every citation points to a source saved in your workspace, and every change is kept as a version you can restore.
How long does it take before wiki pages appear?
Pages are written in the background after each save with AI, usually within minutes. Sources from code, video, social or tool sites are kept out of the wiki by default.
Can I get my wiki out of Pith?
Yes. The whole workspace can be exported at any time as a ZIP with a JSON file and Markdown renderings, and single pages can be shared as links with an expiry date, an optional password and a view limit.
Is the no-code LLM wiki free?
Reading, searching, sharing, exporting and MCP access are free. Saving with AI and wiki pages written by Pith cost a few cents each from a prepaid credit; new accounts get a start credit and a free monthly credit. Current prices are on the pricing page.
Keep reading
Guide · LLM wiki
What is an LLM wiki? Karpathy's pattern, explained
Guide · LLM wiki vs RAG
LLM wiki vs RAG: a decision guide
Guide · Karpathy LLM wiki
Karpathy's LLM wiki: what the gist says, and the open-source builds
Related terms
Try a hosted LLM wiki
Save what you read; Pith writes the cited wiki. No subscription, no credit card to start.