deepaper

One arxiv link → one expert-level deep analysis note, powered by a multi-agent pipeline

English | 中文

deepaper turns arxiv paper links into structured deep analysis notes via a 5-agent pipeline orchestrated by Claude Code. Not summaries -- actionable research notes: 7-section expert analysis with 15 automated quality gates, evidence-based mechanism family trees, cross-domain transfer prescriptions, complete hyperparameter tables and pseudocode, and engineering pitfall warnings for reproducers.

Notes are stored as Obsidian-compatible Markdown + YAML, with Git sync, auto-classification (12 subcategories), powered by Claude Code CLI (Max subscription, no API billing).

How It Works

deepaper runs as a /deepaper slash command inside Claude Code. When you type /deepaper <arxiv-url>, it orchestrates a 5-agent pipeline:

Conductor (Claude Code)
  ├→ [1] Extractor    — reads all pages, outputs structured notes (tables, formulas, related work)
  ├→ [2] Writer-A     — writes frontmatter + executive summary + motivation + methodology      ┐
  ├→ [3] Writer-B     — writes experiments & attribution + critical review                      ├ parallel
  ├→ [4] Writer-C     — writes mechanism transfer analysis + background context                 ┘
  ├→ [5] Critic       — 15 quality gates audit (with inline fallback if API is unavailable)
  └→ [6] Fixer        — patches failed gates (if any)

Each agent has specialized prompts, character-count gates, and automated self-checks. The final output passes 15 quality gates covering factual accuracy, data density, and structural completeness.

deepaper vs Alternatives

Feature	Zotero	Semantic Scholar	Manual Notes	deepaper
Deep Analysis	Bookmarks + highlights	Abstract	Hand-written	7-section multi-agent analysis
Quality Control	None	None	Self-review	15 automated quality gates
Mechanism Mapping	None	None	Yes (time-consuming)	Auto mechanism family tree (ancestors/siblings/descendants)
Citation Tracking	Count only	Citation list	Manual search	Evidence-based descendants (OpenAlex, no API key)
Obsidian Native	None	None	Native	Native (YAML + Markdown + Dataview)
Multi-device Sync	Cloud sync (paid)	None	Git sync (DIY)	Git sync (built-in)
API Cost	None	None	None	None (Max subscription, no per-token billing)

Quick Start

# Install
pip install deepaper

# Initialize project
cd my-papers && deepaper init

# In Claude Code, analyze a paper
/deepaper https://arxiv.org/abs/2512.13961

A markdown note appears in papers/ -- open it directly in Obsidian.

Installation

# Recommended: uv (fast, isolated)
uv tool install deepaper

# Or with pipx
pipx install deepaper

# Or direct install
pip install deepaper

# Install the /deepaper slash command globally
deepaper install

Prerequisite: Claude Code CLI installed and authenticated (Max subscription). PyMuPDF is also needed for PDF text extraction: uv pip install PyMuPDF or pip install PyMuPDF.

Usage

Analyze a Paper (Multi-Agent Pipeline)

In any Claude Code session within a deepaper project:

/deepaper https://arxiv.org/abs/2512.13961

This triggers the full 5-agent pipeline. The Conductor:

1. Downloads the PDF and extracts full text
2. Spawns the Extractor to read all pages and produce structured notes
3. Spawns 3 Writers in parallel to produce the analysis
4. Runs the Critic to verify 15 quality gates
5. Fixes any failed gates
6. Saves to papers/{category}/{title}.md

CLI Commands

# Download a paper PDF + metadata
deepaper download https://arxiv.org/abs/2512.13961

# Save an analysis to the knowledge base
deepaper save 2512.13961 --category llm/pretraining --input /tmp/analysis.md

# Look up citing papers
deepaper cite 2512.13961

# Update an existing note with citation data
deepaper cite --update 2512.13961

# Sync notes to git
deepaper sync

Citation Lookup

deepaper cite 1706.03762

Fetches real citing papers from OpenAlex (free, no API key needed), sorted by citation count. Use --update to inject descendants into an existing note's mechanism family tree.

Git Sync

deepaper sync
deepaper sync --message "Add OLMo 3 analysis"

Auto-runs git pull --rebase, then commits and pushes.

Analysis Output

Each note contains YAML metadata + 7-section deep analysis, verified by 15 quality gates:

Section	Content	Quality Gates
Frontmatter	Baselines (one per line), datasets (with token counts), metrics (with eval config)	Baselines format, metrics config, datasets counts
Executive Summary	TL;DR with specific numbers + old-vs-new comparison + core mechanism	TL;DR contains ≥2 benchmark numbers
Motivation	Baseline pain points with numbers + 3-step causal chain + intuitive analogy	Pain points cite ≥2 baselines, causal chain ≥3 steps
Methodology	Data flow diagram + formulas + numerical walkthrough + pseudocode + hyperparameter tables + design decisions	≥12K chars, design decisions ≥3K chars
Experiments	Full comparison tables (all baselines) + ablation ranking + credibility check	≥2 complete tables, attribution with delta numbers
Critical Review	Hidden costs with numbers + reusable techniques + pitfalls + related work comparison	≥3 hidden costs with numbers
Mechanism Transfer	3-5 primitives + cross-domain prescriptions + family tree (≥4 ancestors, ≥3 siblings)	≥5K chars, prescriptions complete, family counts
Background	External technologies table (≥8 items)	--

Category System

Papers are auto-classified into 12 subcategories:

• LLM: pretraining, alignment, reasoning, efficiency, agent
• RecSys: matching, ranking, llm-as-rec, generative-rec, system
• Multimodal: vlm, generation, understanding
• Other: misc

Obsidian Integration

Open the project root as an Obsidian vault. Use Dataview to query notes:

TABLE date, venue, keywords
FROM "papers"
SORT date DESC
LIMIT 20

Configuration

Parameter	Default	Description
git_remote	--	GitHub/GitLab remote URL (for deepaper sync)
papers_dir	papers	Note storage directory

Citation analysis uses the OpenAlex open API -- no API key required.

For richer influence scoring, optionally configure a Semantic Scholar API key (free) in config.yaml or via SEMANTIC_SCHOLAR_API_KEY env var.

Requirements

• Python 3.10+
• Claude Code CLI (installed and authenticated)
• Max subscription (no API billing)
• PyMuPDF (pip install PyMuPDF) for PDF text extraction
• Git (optional, for deepaper sync)
• Obsidian (optional, for viewing the vault)

License

MIT. See LICENSE.