Surface hidden connections in your Obsidian vault.
Mycelium analyzes what you've been writing about lately — using only local NLP and your git history — and generates an Obsidian canvas that maps the concepts and themes emerging from your recent notes.
Privacy guarantee: All text analysis runs locally. Your notes never leave your machine unless you opt in to LLM synthesis (which sends only concept labels, not your note content, to the Claude API).
- Node.js >= 18
- Your Obsidian vault must be a git repository (at least one commit)
- Optional:
ANTHROPIC_API_KEYfor LLM-powered theme synthesis
git clone https://github.com/lyqed/mycelium
cd mycelium
npm install
npm link # makes `mycelium` available globally
Or run directly with node bin/mycelium.js.
mycelium --vault /path/to/your/vault [options]
| Flag | Default | Description |
|---|---|---|
--vault <path> |
(required) | Path to your Obsidian vault (must be a git repo) |
--since <period> |
7d |
Time period: 7d, 2w, 1m |
--top <n> |
40 |
Number of top concepts to extract |
--dry-run |
false | Print analysis JSON, skip canvas generation |
--no-llm |
false | Skip Claude API, use fallback theme grouping |
--output <path> |
<vault>/_mycelium |
Output directory for canvas file |
# Analyze last 7 days with LLM synthesis
mycelium --vault ~/Documents/MyVault
# Two weeks, no LLM, dry run to inspect concepts
mycelium --vault ~/Documents/MyVault --since 2w --no-llm --dry-run
# Last month, custom output directory
mycelium --vault ~/Documents/MyVault --since 1m --output ~/Desktop/maps
The canvas is written to <vault>/_mycelium/mycelium-YYYY-MM-DD.canvas. Open it in Obsidian with the Canvas core plugin enabled.
- Git extraction — reads lines added to
.mdfiles in the given period usinggit log - NLP pipeline — cleans markdown, tokenizes, removes English + Hebrew stopwords, stems with Porter stemmer, builds unigrams/bigrams/trigrams with n-gram absorption
- Scoring —
frequency × log2(fileSpread + 1)rewards concepts that appear across many files - Theme synthesis — sends concept labels (not note content) to Claude Haiku; falls back to alphabetical grouping if unavailable
- Canvas generation — produces Obsidian
.canvasJSON with a deterministic 2-column grid layout, color-coded by concept score
- Hebrew NLP (stemming, tokenization aware of nikud)
- Recurring concept tracking across multiple canvas snapshots
- Concept diff: highlight new vs. returning concepts
- Interactive concept filtering via Obsidian plugin
- Local LLM support (Ollama)