Dump

An Obsidian plugin that turns dumped URLs, podcasts, and text into a cross-linked wiki maintained by an AI agent.

Based on Andrej Karpathy's LLM Wiki idea: instead of bookmarking and searching, have an AI maintain a living knowledge base that synthesizes everything you capture.

Philosophy

Bookmarking is a graveyard. You save articles, podcasts, and notes "for later" and never read them again. Search is barely better — you have to remember the right keywords years after the fact.

Dump flips it: every capture goes through an AI agent that reads the source, distills the key ideas, drops them into wiki pages, and cross-links them with everything else you've ever dumped. The wiki gets smarter every time you feed it. You ask questions of your own knowledge base and get answers grounded in your sources.

How it works

1. You dump a URL, a podcast link, or some text.
2. The plugin extracts the content (Readability for articles, Whisper for podcasts).
3. An AI agent saves the raw source, creates or updates wiki pages, and cross-references them.
4. The index and log are updated automatically.

Commands

Command	Description
Dump URL	Open a modal to paste a URL or text
Dump clipboard	Grab clipboard content and ingest
Dump selection	Dump selected text from the active note
Query knowledge base	Search and synthesize answers from dumped content
Lint wiki	Audit for orphans, broken links, and contradictions
Process inbox	Manually process queued inbox files
Show status	Display KB stats (sources, pages, categories)

Vault structure

Dump/
├── inbox/          # New captures, awaiting processing
├── raw/            # Immutable source captures (YYYY-MM-DD-slug.md)
│   └── assets/     # Images, screenshots
├── wiki/           # AI-maintained wiki pages (slug.md)
│   ├── index.md    # Master index by category
│   └── log.md      # Chronological operation log
└── outputs/        # Reports, query answers, lint results

Raw files are immutable. Wiki pages are living documents that grow as you add more sources. Every claim cites its source.

Install

Via BRAT (recommended for now)

1. Install BRAT from Community Plugins
2. In BRAT settings, click "Add Beta plugin"
3. Enter: omarshahine/obsidian-dump-llm-wiki
4. Enable the plugin in Community Plugins

Manual

1. Download main.js, manifest.json, and styles.css from the latest release
2. Create <vault>/.obsidian/plugins/dump-llm-wiki/
3. Copy the files there
4. Enable in Community Plugins

Setup

Pick one of two modes — they're not mutually exclusive, you can configure both and the plugin will fall back from CLI to API if needed.

CLI mode (recommended)

If you have Claude Code installed, the plugin will shell out to it for processing. This is the default and the most powerful option — Claude Code can use Read, Write, Edit, Glob, Grep, Bash, and WebFetch to maintain the wiki.

1. Install Claude Code
2. In Settings > Dump, set "CLI tool" to Claude
3. Pick a model (Opus, Sonnet, or Haiku)

No API key required. The plugin auto-detects the CLI on macOS, including Homebrew and nvm install paths.

API mode

If you'd rather use the Anthropic or OpenAI API directly:

1. Open Settings > Dump
2. Pick Anthropic or OpenAI
3. Enter your API key (stored per-provider in your OS keychain via SecretStorage, never synced)
4. Click the refresh icon next to Model to load available models from the API

Optional: podcast transcription

To dump podcasts, add an OpenAI API key under "Transcription". Whisper or gpt-4o-transcribe will transcribe the audio before ingest.

Settings

Setting	What it does
CLI tool	Claude Code is the primary processor when installed
CLI model	Opus, Sonnet, or Haiku — passed to Claude Code
Provider	Anthropic or OpenAI for API fallback
API key	Stored in OS keychain, never in data.json
Dump folder	Folder name within your vault (default: Dump)
Auto-lint interval	Disabled, daily, weekly, or monthly health checks
Max files per session	Safety cap to prevent runaway processing costs
Auto-process delay	Wait after Obsidian boot before processing the inbox
Whisper model	whisper-1 or gpt-4o-transcribe
Max tool iterations	Safety limit for the API agent loop

Design decisions

Decision	Why
Inbox/raw split	Captures land in inbox/ instantly, then move to raw/ after processing. Raw is immutable; the inbox is your queue.
CLI before API	Claude Code has more capabilities (Bash, WebFetch with extraction tools) and lower latency. API is the fallback.
Per-provider API keys	Stored in OS keychain via Obsidian's SecretStorage. Never synced, never in data.json.
Two API providers	Anthropic and OpenAI cover the best models. No bloat.
Session caps	Prevents one runaway dump from torching your API budget.
Desktop only	The CLI integration uses Node child_process, which isn't available on mobile.

Development

git clone https://github.com/omarshahine/obsidian-dump-llm-wiki.git
cd obsidian-dump-llm-wiki
npm install
npm run dev    # Watch mode
npm run build  # Production build

Symlink into your vault for testing:

ln -s /path/to/obsidian-dump-llm-wiki /path/to/vault/.obsidian/plugins/dump-llm-wiki

Credits

Inspired by Andrej Karpathy's LLM Wiki concept.

License

MIT