Skillget

paper7

mcp
Security Audit
Fail
Health Warn
  • No license — Repository has no license file
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 9 GitHub stars
Code Fail
  • rm -rf — Recursive force deletion command in install.sh
  • rm -rf — Recursive force deletion command in paper7.sh
  • rm -rf — Recursive force deletion command in tests/test_browse.sh
  • rm -rf — Recursive force deletion command in tests/test_doi.sh
  • rm -rf — Recursive force deletion command in tests/test_get_modes.sh
  • rm -rf — Recursive force deletion command in tests/test_pubmed_get.sh
  • rm -rf — Recursive force deletion command in tests/test_s2.sh
Permissions Pass
  • Permissions — No dangerous permissions requested
Purpose
This tool converts arXiv and PubMed academic papers into clean Markdown formatted for LLMs, acting as an MCP server to help AI agents search and fetch literature directly.

Security Assessment
Risk: Medium. The tool inherently makes outbound network requests to academic APIs (arXiv, PubMed, Semantic Scholar) and requires an internet connection to function. Because it is written in Bash, it relies heavily on executing shell commands. While no dangerous system permissions or hardcoded secrets were found, the codebase contains multiple `rm -rf` recursive force deletion commands across the installation script, main script, and test files. If these deletion paths are manipulated or behave unexpectedly, they pose a risk of unintended local data loss.

Quality Assessment
The repository is actively maintained, with its most recent push occurring today. However, the project suffers from low community visibility, having accrued only 9 GitHub stars. More importantly, the project lacks a designated open-source license. Without a license, the code is technically proprietary, meaning developers do not have explicit legal permission to use, modify, or distribute it.

Verdict
Use with caution: while the utility is actively maintained and fills a specific LLM need, developers should be wary of the unlicensed codebase and carefully inspect the aggressive shell deletion commands before running or integrating it.
SUMMARY

arXiv papers as clean Markdown for LLMs. 97% smaller than PDF. Zero dependencies. Bash-only.

README.md

paper7

Turn any arXiv or PubMed paper into clean Markdown — at runtime, with zero dependencies.
97% smaller than PDF. 86% smaller than raw HTML.

See the benchmark →

Install

curl -sSL https://raw.githubusercontent.com/lucianfialho/paper7/main/install.sh | bash

AI Agent Skill

paper7 ships with an installable skill that teaches AI coding agents (Claude Code, Cursor, Codex) how to use it. Once installed, your agent can search arXiv or PubMed and fetch papers directly into the conversation.

# Core skill — search and fetch papers
npx skills add lucianfialho/paper7 --skill paper7

# Research skill — guided literature review before implementation
npx skills add lucianfialho/paper7 --skill paper7-research

Or manually as a Claude Code slash command:

mkdir -p ~/.claude/commands
curl -sL https://raw.githubusercontent.com/lucianfialho/paper7/main/claude-code/paper7.md \
  -o ~/.claude/commands/paper7.md

After installing, try prompts like:

  • paper7 search "attention mechanism" — search arXiv
  • paper7 get 2401.04088 — fetch a compact paper header with summary + section index
  • paper7 get 2401.04088 --detailed --range 35:67 — fetch just one detailed slice
  • "Research whether LoRA or full fine-tuning is better for my use case" — triggers the research skill

Usage

# Search arXiv (default)
paper7 search "mixture of experts" --max 5

# Search PubMed (biomedical, clinical, pharmacological)
paper7 search "psilocybin hypertension" --source pubmed --max 5

# Fetch a paper (compact indexed header by default)
paper7 get 2401.04088                          # arXiv
paper7 get 2401.04088 --detailed               # full paper
paper7 get 2401.04088 --detailed --range 35:67 # just lines 35-67 from full paper
paper7 get https://arxiv.org/abs/2401.04088
paper7 get pmid:38903003                       # PubMed (abstract only)
paper7 get doi:10.1101/2023.12.15.571821       # any DOI via Crossref (bioRxiv, medRxiv, etc.)
paper7 get 2401.04088 --no-refs                # strip references
paper7 get 2401.04088 --no-cache               # force re-download

# Find source code
paper7 repo 2401.04088

# List references via Semantic Scholar (requires jq)
paper7 refs 1706.03762 --max 5
paper7 refs 1706.03762 --json | jq '.data | length'   # pipe raw JSON

# Manage your local knowledge base
paper7 list                                     # show cached papers
paper7 cache clear 2401.04088                   # remove one
paper7 cache clear                              # clear all

# Export to an Obsidian-compatible vault (frontmatter + wikilinks)
paper7 vault init ~/Documents/ArxivVault        # configure vault path once
paper7 vault 2401.04088                         # export one paper
paper7 vault all                                # export every cached paper

# Browse the local cache interactively (requires fzf; glow recommended)
paper7 browse                                   # fzf picker + preview; Enter renders, Esc quits

# Pipe to anything
paper7 get 2401.04088 | claude "which section should I read?"      # compact header first
paper7 get 2401.04088 --detailed --range 35:67 | claude "explain"  # just one section slice
paper7 get 2401.04088 --detailed | llm "summarize"                 # simon willison's llm
paper7 get 2401.04088 --detailed | pbcopy                          # clipboard (macOS)
paper7 get 2401.04088 --detailed --no-refs > paper.md              # save full paper to file

# End-to-end: search PubMed, then fetch and summarize
paper7 search "psilocybin hypertension" --source pubmed --max 3
paper7 get pmid:38903003 | claude "summarize the clinical case"

Sources

paper7 queries one of two sources per search; pick based on the topic:

  • arxiv (default): physics, computer science, machine learning, math, quantitative biology. Full-text available.
  • pubmed: biomedical, clinical, and pharmacological literature. Abstracts only (full text on PMC is a separate pipeline).

PubMed results use a pmid: prefix on the ID; arXiv IDs keep the native YYMM.NNNNN form. Both coexist in the same local cache.

paper7 get doi:<DOI> covers anything with a DOI — bioRxiv, medRxiv, PsyArXiv, ChemRxiv, journal articles — via Crossref (metadata + abstract). bioRxiv/medRxiv full text isn't available (their pages block direct HTTP); the rendered Markdown includes a **Full text:** link.

Semantic Scholar is also wired in as a metadata layer (not a full-paper fetcher): paper7 refs <id> lists canonical references, and paper7 get --detailed enriches its Markdown header with an auto-generated **TLDR:** line when one exists. Plain paper7 get emits a compact header with a summary and line-indexed sections so agents can fetch only the ranges they need.

For per-source endpoints, rate limits, auth, and known gaps see docs/sources.md.

Benchmark

Tested with 5 landmark papers (Attention, RAG, Mixtral, GPT-4, LoRA) — 169 pages total:

                          Size (5 papers combined)

  Raw PDF       ████████████████████████████████████████████████  12,140KB
  HTML (ar5iv)  ██████████████████                                 2,522KB
  paper7        ██                                                   349KB  (-97% vs PDF)
Paper Pages PDF HTML paper7 vs PDF vs HTML
Attention Is All You Need 15 2,163KB 343KB 40KB -98% -88%
RAG 12 864KB 301KB 68KB -92% -77%
Mixtral of Experts 16 2,417KB 216KB 31KB -98% -85%
GPT-4 Technical Report 100 5,122KB 635KB 116KB -97% -81%
LoRA 26 1,571KB 1,024KB 91KB -94% -91%
Total 169 12,140KB 2,522KB 349KB -97% -86%

Reproduce with ./benchmark/run.sh.

How it works

arXiv flow (full-text):

  1. Search arXiv API for papers by keyword
  2. Fetch full text from ar5iv (HTML version of arXiv — no PDF parsing)
  3. Convert HTML to clean Markdown with proper ## headers, paragraphs, and structure
  4. Cache locally at ~/.paper7/cache/<arxiv_id>/

PubMed flow (abstract-only):

  1. Search NCBI E-utilities (esearch + esummary) for papers by keyword
  2. Fetch abstract via efetch (XML) — full text lives on PMC and is a separate pipeline
  3. Convert XML to clean Markdown with title, authors, journal, DOI, and abstract (labeled sections preserved)
  4. Cache locally at ~/.paper7/cache/pmid-<NNN>/

paper7 skips PDF parsing entirely. For arXiv, ar5iv provides the same content as HTML without binary layout overhead. For PubMed, the E-utilities XML is already structured metadata. In both cases, paper7 extracts the body, converts tags to Markdown, and strips everything else.

Why not just use PDF?

Raw PDF paper7
Size ~12MB for 5 papers ~350KB (-97%)
Structure Flat binary, no sections Markdown with ## headers
Two-column layout Broken text flow Linear reading order
Page headers/footers Repeated every page Removed
Math notation Garbled or requires Vision API Cleaned
Metadata Mixed into body text Structured header
Local cache No Built-in knowledge base
Dependencies Vision API or poppler curl

Research

The clean-text-over-raw-PDF approach is backed by academic research. See examples/research-kb/ for a knowledge base built with paper7 itself:

  • Lost in the Middle (Liu et al., 2023) — LLMs lose 20%+ performance when relevant info is buried in long, noisy contexts
  • PDF-WuKong (Xie et al., 2024) — sparse sampling reduces tokens by ~89% while improving comprehension
  • Comparative Study of PDF Parsing (Adhikari & Agarwal, 2024) — recommends Markdown/LaTeX for scientific documents

CLI reference

paper7 <command> [options]

Commands:
  search <query>       Search papers by keyword (arXiv or PubMed)
  get <id>             Fetch paper; compact header by default, full text with --detailed
                       id shapes: arXiv (YYMM.NNNNN), pmid:NNN, doi:10.XXXX/...
  refs <id>            List references via Semantic Scholar (requires jq)
                       id shapes: YYMM.NNNNN (arXiv),
                                  https://arxiv.org/abs/... (arXiv URL),
                                  pmid:NNNNN (PubMed abstract)
  repo <id>            Find GitHub repositories for an arXiv paper
  list                 List cached papers (arXiv + PubMed)
  cache clear [id]     Clear cache (all, or a specific arXiv/pmid id)
  vault init <path>    Configure Obsidian-compatible vault
  vault <id>|all       Export arXiv paper(s) to vault with frontmatter + wikilinks
  browse               Interactive fzf picker over the local cache (glow renderer)

Options:
  --source SOURCE      search only — arxiv (default) or pubmed
  --max N              Max search results / references (default: 10)
  --sort relevance|date  Sort search results
  --no-refs            Strip references section (arXiv only; no-op for PubMed)
  --no-cache           Force re-download
  --no-tldr            Skip Semantic Scholar TLDR enrichment in `get`
  --detailed           Emit the full paper instead of the compact indexed header
  --range START:END    Detailed-only line slice from the full paper
  --json               Emit raw JSON (refs only)
  --help, -h           Show help
  --version, -v        Show version

License

MIT

Like context7 but for academic papers. Pure Bash — requires only curl, sed, grep, awk.

Reviews (0)

No results found