Open Source Implementation of Karpathy's LLM Wiki. Upload documents, connect your Claude account via MCP, and have it write your wiki !
# Add to your Claude Code skills
git clone https://github.com/lucasastorian/llmwikiLast scanned: 5/5/2026
{
"issues": [],
"status": "PASSED",
"scannedAt": "2026-05-05T06:19:43.117Z",
"semgrepRan": false,
"npmAuditRan": true,
"pipAuditRan": true
}No comments yet. Be the first to share your thoughts!
Based on votes and bookmarks from developers who liked this skill
Open-source implementation of Karpathy's LLM Wiki (spec).
I built this because research folders accumulate useful material faster than I can keep summaries, links, and citations current by hand. LLM Wiki offloads that editing work to Claude so I can focus on source selection and analysis instead.
Point it at a folder, start the local app, and connect Claude over MCP. From there, Claude reads your sources, writes wiki pages, and keeps links and citations in sync.

wiki/, maintains cross-references and footnote citations.Requirements: Python 3.11+, Node.js 20+
git clone https://github.com/lucasastorian/llmwiki.git
cd llmwiki
# Install Python deps
cd api && python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cd ..
# Install web deps
cd web && npm install && cd ..
# Initialize a workspace (point at any folder with your files)
./llmwiki init ~/research
# Start the app
./llmwiki serve ~/research
Open localhost:3000. Your files are indexed, wiki is scaffolded, ready to go.
./llmwiki mcp-config ~/research
This prints a JSON snippet for claude_desktop_config.json (Claude Desktop) or .claude/settings.json (Claude Code). One workspace runs as one MCP server entry, so if you have multiple research folders, add one entry per folder.
Then tell Claude: "Read the guide, then ingest my sources and start building the wiki."
./llmwiki open ~/research
Does everything: init if needed, start servers, open browser, print MCP config hint.
| Command | What it does |
|---------|-------------|
| llmwiki open <folder> | Init + serve + open browser |
| llmwiki init <folder> | Create .llmwiki/ + wiki/, index existing files |
| llmwiki serve <folder> | Start API on :8000 + web on :3000 |
| llmwiki mcp <folder> | Run stdio MCP server (for Claude config) |
| llmwiki mcp-config <folder> | Print claude_desktop_config.json snippet |
| llmwiki reindex <folder> | Rebuild the index from disk |
LLM Wiki adds two things to your folder. Source files are not moved or modified.
~/research/ # Your existing files (untouched)
papers/paper.pdf
notes.md
data.xlsx
wiki/ # Generated pages (created by LLM Wiki)
overview.md
log.md
concepts/
attention.md
.llmwiki/ # Index + cache (hidden, rebuildable)
index.db
cache/
wiki/ — ordinary markdown files. Edit them in any editor. Claude writes and updates them via MCP..llmwiki/ — SQLite search index and processed artifacts. Delete it anytime; llmwiki reindex rebuilds from the source files.By default, indexing, storage, and file writes happen on your machine. No cloud services required.
Once connected, Claude has these tools:
| Tool | Description |
|------|-------------|
| guide | Explains how the wiki works, lists what's in the workspace |
| search | Browse files (list) or full-text search (search) |
| read | Read documents — PDFs with page ranges, glob batch reads |
| write | Create wiki pages, edit with str_replace, append. SVG/CSV assets |
| delete | Delete documents by path or glob pattern |
All writes go to disk first, then update the search index. If Claude creates /wiki/concepts/attention.md, that file appears on disk immediately.
┌──────────────┐ ┌──────────────┐ ┌──────────────┐
│ Next.js │────▶│ FastAPI │────▶│ SQLite │
│ Frontend │ │ Backend │ │ (local) │
└──────────────┘ └──────┬───────┘ └──────────────┘
│
┌──────┴───────┐
│ MCP Server │◀──── Claude Desktop / Code
│ (stdio) │
└──────────────┘
│
┌──────┴───────┐
│ Filesystem │ ← source of truth
└──────────────┘
The filesystem is the source of truth. SQLite is a derived index — it accelerates search and stores extracted page data, but it can always be rebuilt from the files. A background file watcher picks up changes you make outside the app.
All processing runs locally. No API keys required for basic usage.
| Format | Parser | Notes | |--------|--------|-------| | PDF | pdf-oxide | Rust-based text extraction. Works well for text-heavy papers. Scanned PDFs still benefit from real OCR. | | Markdown/Text | native | Indexed and chunked directly | | HTML | webmd | Strips nav/ads, extracts clean markdown | | Excel/CSV | openpyxl | Sheet-by-sheet extraction | | Images | native | Stored as-is, viewable inline | | Word/PowerPoint | LibreOffice | Optional. Install LibreOffice for office conversion; without it, these formats are stored but not extracted. |
Set MISTRAL_API_KEY for higher-quality PDF OCR with better table and layout detection. pdf-oxide is the free default and handles most text-heavy documents well enough.
If you want to run the hosted version (like llmwiki.app) with Postgres, Supabase auth, and S3:
psql $DATABASE_URL -f supabase/migrations/001_initial.sql
cd api
pip install -r requirements.txt
MODE=hosted DATABASE_URL=postgresql://... uvicorn main:app --port 8000
cd mcp
pip install -r requirements.txt
MODE=hosted DATABASE_URL=postgresql://... uvicorn server:app --port 8080
cd web
npm install
NEXT_PUBLIC_MODE=hosted \
NEXT_PUBLIC_SUPABASE_URL=https://your-ref.supabase.co \
NEXT_PUBLIC_SUPABASE_ANON_KEY=your-anon-key \
NEXT_PUBLIC_API_URL=http://localhost:8000 \
npm run dev
API
MODE=hosted
DATABASE_URL=postgresql://...
SUPABASE_URL=https://your-ref.supabase.co
AWS_ACCESS_KEY_ID=...
AWS_SECRET_ACCESS_KEY=...
S3_BUCKET=your-bucket
MISTRAL_API_KEY= # optional, for better PDF OCR
CONVERTER_URL= # optional, for office conversion
Web
NEXT_PUBLIC_MODE=hosted
NEXT_PUBLIC_SUPABASE_URL=https://your-ref.supabase.co
NEXT_PUBLIC_SUPABASE_ANON_KEY=your-anon-key
NEXT_PUBLIC_API_URL=http://localhost:8000
Personal wikis usually fail on maintenance, not intent. Someone has to update links, fix stale summaries, merge overlapping pages, and keep citations aligned with the source material. That work scales with the number of sources, and people stop doing it.
LLM Wiki offloads that editing work. You choose the source material and direct the analysis. Claude handles the repetitive bookkeeping — updating cross-references, keeping summaries current, flagging contradictions, touching the 15 pages that a single new source affects.
Apache 2.0