Literature Search Arxiv

arXiv Search and Retrieval

Prerequisites

1. uv: Read the uv skill and follow its Setup instructions to ensure

uv is installed and on PATH.

1. User Notification: If LICENSE_NOTIFICATION.txt does not already exist in

this skill directory then (1) prominently notify the user to check the terms at https://info.arxiv.org/help/api/index.html and to always check the license of the papers retrieved by the skill for any restrictions, then (2) create the file recording the notification text and timestamp.

Core Rules

• Terms of Use: You MUST respect arXiv's Terms of Use.
• Maximum 1 request every 3 seconds.
• The provided utility scripts handle rate limiting automatically. Always

use these scripts rather than writing your own curl/python requests.

• If this skill is used, ensure this is mentioned in the output AND list the

URLs of all papers that were used in producing the output.

Utility Scripts

1. Search and Extract Metadata

Search arXiv and return a clean JSON array of matching papers.

uv run scripts/search_arxiv.py --query "au:einstein AND ti:relativity" \
  --max_results 5 2>/dev/null > /tmp/arxiv_search_results.json

> Important: The tool outputs a large JSON result to stdout. Requesting 100+ > results will produce a massive JSON that might exceed your context length. > Limit --max_results (e.g., 5-10) or paginate carefully using --start. > Always redirect output to a file and parse it separately, otherwise terminal > output will be truncated.

Returned Metadata: JSON results include id, title, summary, published, authors, pdf_url, primary_category, doi, journal_ref, and comment. Note: the doi field only contains DOI information in case the paper has an external DOI and if only an arXiv-issued DOI exists, this is DOI is not returned.

Options:

• --query: Search string. See

[references/querysyntax.md](references/querysyntax.md) for advanced syntax.

• --id_list: Comma-separated list of arXiv IDs to fetch directly (e.g.,

1706.03762v5).

• --start: Pagination offset (default 0).
• --max_results: Number of results to return (default 10).
• --sort_by: relevance, lastUpdatedDate, or submittedDate. (Use

--sort_by submittedDate --sort_order descending for the most recent papers).

• --sort_order: ascending or descending.

2. Download Paper (PDF or HTML)

Download the full text of a paper to your local workspace for reading.

uv run scripts/download_paper.py --id 1706.03762 --format pdf --output attention.pdf

Options:

• --id: The arXiv ID (e.g., 1706.03762 or 1706.03762v5).
• --format: pdf or html. Note: HTML is only available for newer papers.
• --output: Filepath to save the downloaded document.

> Important: when downloading papers, make sure you download them to a > location where you do not overwrite other files and do not clutter existing > directory structure.

3. Download Paper Source (tar.gz)

Download the LaTeX source files of a paper to your local workspace. Note that not all papers have source available.

uv run scripts/download_paper_source.py --id 2010.11645 --output source.tar.gz

Options:

• --id: The arXiv ID (e.g., 2010.11645).
• --output: Filepath to save the downloaded tar.gz file.

> Caution: Care should be exercised when untar'ing the downloaded file for > security and to avoid cluttering your filesystem, as archives may contain many > files or unexpected directory structures. > > Safe Extraction Requirements: NEVER extract directly into your working > directory! Always extract into a dedicated new directory: bash mkdir > paper_source && tar -xzf source.tar.gz -C paper_source

Reference

• Advanced Query Syntax: See

[references/querysyntax.md](references/querysyntax.md) for prefixes (au, ti, abs), booleans, and date filtering.

Workflow

1. Search for papers using search_arxiv.py. Review the JSON summaries.
2. If full text is needed, use download_paper.py to fetch the PDF or HTML.
3. If downloading a PDF, verify the PDF is not empty or corrupted.
4. Read the downloaded file using standard file reading tools.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: google-deepmind
• Source: google-deepmind/science-skills
• License: Apache-2.0
• Homepage: https://antigravity.google/use-cases/science

Install and usage instructions live in the source repository linked above.