A Claude Code Plugin for Self-maintaining research knowledge graph for Claude Code + Obsidian.


_{Knowledge graph view — 🟡 papers · 🟢 findings · 🟣 fields · 🔴 authors. Example shown: Typhoon.AI research papers.}	_{Per-paper info page — 4-section summary (incl. critique), atomic findings, metadata + quality scores, and typed edges to related work.}

Keep every paper you care about. See how they connect. Drop in a URL, arXiv ID, DOI, or PDF — Claude files it into an Obsidian vault, extracts the atomic claims, and wires them to everything you've read before with typed edges: supports, contradicts, extends, uses, similar-to.

Based on Andrej Karpathy's LLM Wiki pattern, tuned for research papers. 4 LLM calls per paper. Zero manual filing. Linking cost stays constant as the graph grows.

Features

4-section triage summary per paper — Key Takeaways → Background → Main Idea & Summary → Critique. The critique is generated by a reasoning model, so you get critical thinking, not a polite restatement of the abstract.
Structured metadata + quality scores — authors, venue, date, fields, and a quality assessment (credibility, experimental rigor, reproducibility) on every paper.
Atomic findings as graph nodes — one claim per file, with frontmatter and backlinks.
Typed edges — supports, contradicts, extends, uses, similar-to; auto-aggregated to paper-level (cites, builds-on, …).
Deterministic citation matching — arXiv ID → DOI → fuzzy title → first-author+year. The LLM never sees the reference list.
Constant-cost linking — pre-filter caps the LLM at ≤30 candidate findings, so the 100th paper costs the same as the 10th.
Cross-paper contradiction surfacing — bidirectional, with a pre-built Dataview view.
Pre-built Dataview views — by-author, by-field, contradictions, high-credibility, recent-papers.
Obsidian-native — plain wikilinks + YAML frontmatter, no custom app, no lock-in.
Local model support — point at Ollama via ANTHROPIC_BASE_URL.

Quick Start

# 1. Install the plugin
claude --plugin-dir /path/to/claude-paperloom

# 2. Set up Python deps (plugin expects a venv at the plugin root)
cd /path/to/claude-paperloom
python3.11 -m venv .venv
.venv/bin/pip install pyyaml pymupdf

# 3. Scaffold your vault (default ~/PaperLoom)
/paperloom:init

# 4. Ingest your first paper
/paperloom:ingest https://arxiv.org/abs/1706.03762

Open the vault folder in Obsidian and turn off Restricted Mode once to activate Dataview.

Commands

You run	Claude does
`/paperloom:init [path]`	Scaffold the vault. Idempotent. Seeds Dataview.
`/paperloom:ingest <url\|arxiv-id\|doi\|pdf>`	Fetch, summarize, extract findings, link. Skips if already in vault.
`/paperloom:query <question>`	Search + synthesize across the vault with wikilink citations.
`/paperloom:lint`	Health check. Auto-wires near-duplicate findings with bidirectional `similar-to`.

Docs

Pipeline — the 4-LLM-call ingest pipeline, section detection, models, token cost.
Linking — five-layer linking, relation types, paper-level aggregation.
Architecture — vault layout, plugin structure, design principles, out-of-scope.

Features

Quick Start

Commands

Docs

Reviews (0)