arxiv

name	arxiv
description	Search arXiv scholarly papers by keyword, title, author, abstract, category, or arXiv ID using curl + xmllint through the bash tool. Wraps the official arXiv API (export.arxiv.org/api/query), which returns Atom XML — parsed with xmllint local-name() XPath, no jq and no extra Python dependencies. Covers fielded search (all:/ti:/au:/abs:/cat:), boolean queries (AND/OR/ANDNOT), lookup by id_list, sorting by relevance/date, and paging. Needs curl + xmllint installed; no API key required. Use when the user wants to find, look up, or list academic papers / preprints on arXiv.
category	research

arXiv paper search (搜 arXiv 论文)

Search and look up scholarly papers on arXiv entirely with curl + xmllint through the bash tool. No dedicated Python tool and no extra Python dependencies — it is shell only. No API key is required.

The arXiv API returns Atom XML (not JSON), so this skill parses with xmllint using local-name() XPath. jq is not used.

Reply in Chinese unless the user clearly uses another language.

Prerequisites

• curl and xmllint installed. Check with command -v curl xmllint. On Windows the bundled msys64 / an Anaconda install provides both; xmllint ships with libxml2 (apt-get install libxml2-utils / brew install libxml2 if missing).

command -v curl xmllint >/dev/null || { echo "need curl + xmllint"; exit 1; }
API=https://export.arxiv.org/api/query

Always use https:// — plain http://export.arxiv.org returns a 301 redirect with an empty body and no results.

Never fabricate arXiv IDs, titles, or URLs (hard rule)

Every arXiv id (e.g. 2101.00001), arxiv.org/abs/... page link, arxiv.org/pdf/... file URL, paper title, and author name you give the user must come from an actual API response you just received — copy it verbatim from the curl | xmllint output. Do not guess, invent, or hand-craft an arXiv id, a title, or a link from memory. A made-up id will resolve to the wrong paper or 404 and is worse than admitting you couldn't search.

If you can't run a real query — curl/xmllint missing, network blocked, or the feed returned zero entry elements — then say so plainly and stop. Report the exact blocker (which dependency, the HTTP status, or that totalResults/entry count was 0) and what the user should try (install the tool, refine the query). Never substitute a fabricated citation in place of a real search result.

Search

Build every request with curl -sG + --data-urlencode so spaces and non-ASCII terms are encoded safely.

# Keyword search across all fields, newest first.
curl -sG "$API" \
  --data-urlencode "search_query=all:large language models" \
  --data-urlencode "start=0" \
  --data-urlencode "max_results=10" \
  --data-urlencode "sortBy=submittedDate" \
  --data-urlencode "sortOrder=descending" \
  -o /tmp/arxiv.xml

Query fields (prefix inside search_query)

• all: — anywhere (title, abstract, authors, comments…)
• ti: — title, abs: — abstract, au: — author, co: — comment
• cat: — arXiv category, e.g. cat:cs.AI, cat:hep-th, cat:stat.ML
• jr: — journal ref, rn: — report number

Boolean queries (use spaces, never a literal +)

Combine terms with AND, OR, ANDNOT, separated by spaces — --data-urlencode turns spaces into %20, which arXiv accepts. Do not type a literal + between terms: it gets encoded to %2B and silently returns zero results.

# Papers in cs.AI by an author named Bengio:
--data-urlencode "search_query=cat:cs.AI AND au:Bengio"

# Group with parentheses and quote exact phrases with %22 ... %22:
--data-urlencode 'search_query=ti:%22retrieval augmented%22 ANDNOT cat:cs.CL'

Parameters (query string)

• search_query — the fielded query above. Either search_query or id_list (or both) must be present.
• start (default 0) and max_results (default 10) — paging. Keep max_results modest (≤ 100 for interactive use; the API allows larger with start paging but throttles).
• sortBy — relevance (default), lastUpdatedDate, or submittedDate.
• sortOrder — ascending or descending.

Look up specific papers by ID

# One or more arXiv ids (comma-separated), no search_query needed:
curl -sG "$API" --data-urlencode "id_list=2101.00001,1706.03762" -o /tmp/arxiv.xml

Accepts both new-style ids (2101.00001, optionally v2) and old-style (cond-mat/0011267). Combine id_list with search_query to filter within a set.

Parse the Atom XML (xmllint local-name recipe)

The feed uses a default namespace, so bind nothing and match by local-name(). Pull opensearch:totalResults first, then iterate entries by index so fields from different entries never get flattened together.

X=/tmp/arxiv.xml
total=$(xmllint --xpath 'string(//*[local-name()="totalResults"])' "$X" 2>/dev/null)
n=$(xmllint --xpath 'count(//*[local-name()="entry"])' "$X" 2>/dev/null)
echo "total matches on arXiv: ${total:-0}; entries in this page: ${n:-0}"

for i in $(seq 1 "${n:-0}"); do
  e="(//*[local-name()=\"entry\"])[$i]"
  id=$(xmllint --xpath "string($e/*[local-name()=\"id\"])" "$X")
  title=$(xmllint --xpath "string($e/*[local-name()=\"title\"])" "$X" | tr -s ' \n' ' ')
  cat=$(xmllint --xpath "string($e/*[local-name()=\"primary_category\"]/@term)" "$X")
  pub=$(xmllint --xpath "string($e/*[local-name()=\"published\"])" "$X")
  pdf=$(xmllint --xpath "string($e/*[local-name()=\"link\"][@title=\"pdf\"]/@href)" "$X")
  authors=$(xmllint --xpath "$e/*[local-name()=\"author\"]/*[local-name()=\"name\"]/text()" "$X" 2>/dev/null | paste -sd'; ')
  printf '%s\n  %s\n  %s | %s | %s\n  pdf: %s\n\n' "$title" "$authors" "$cat" "${pub%%T*}" "$id" "$pdf"
done

Useful per-entry fields: id (abs URL — the trailing token is the arXiv id), title, summary (abstract), published / updated, author/name, primary_category/@term, every category/@term, and links (@title="pdf" for the PDF, rel="alternate" for the HTML page). Trim whitespace in title/summary with tr -s ' \n' ' ' — arXiv wraps those fields across lines.

Rate limits & etiquette (important)

arXiv asks callers to wait at least 3 seconds between requests and avoid bursts. When paging or issuing several queries, sleep 3 between calls. A single search is fine; do not hammer the endpoint in a loop.

Present results

Give the user the title, authors, primary category, date, arXiv id, and links for each hit — all copied verbatim from the parsed output. Link the abstract page (https://arxiv.org/abs/<id>) and the PDF. If the user wants the file delivered, download the PDF and hand it off with a [SEND:<absolute-path>] line, the same way other skills deliver output files:

curl -sL "https://arxiv.org/pdf/<id>" -o "/tmp/<id>.pdf"

If totalResults is 0 or the query looks too narrow, say so and suggest broadening the terms or dropping a field prefix — never invent a paper to fill the gap.