AGENT SKILL

search-arxiv

Search arXiv by paper ID or keyword and return title, abstract, authors, and BibTeX. Use when asked to look up an arXiv paper, fetch metadata for a citation, check what a paper is about, generate BibTeX for an arXiv ID, or find papers on a topic.

License 许可
GPL-3.0
最近更新
2026-11-13
Stars
★ 0
Forks
0

01 前言 Front Matter

name
search-arxiv
description
Search arXiv by paper ID or keyword and return title, abstract, authors, and BibTeX. Use when asked to look up an arXiv paper, fetch metadata for a citation, check what a paper is about, generate BibTeX for an arXiv ID, or find papers on a topic.

03 SKILL.md 正文

search-arxiv

Use the Atom API for metadata; optional arxiv-mcp-server adds full-paper access (see §MCP enhancement).


Step 1 — Identify the query type

  • arXiv ID (e.g. 2407.19122, math/0603460, 2407.19122v2): go to Step 2a.
  • Keyword / author / title: go to Step 2b.

Step 2a — Fetch by arXiv ID

WebFetch the Atom endpoint:

https://export.arxiv.org/api/query?id_list=<ID>

Example: https://export.arxiv.org/api/query?id_list=2407.19122

Prompt: "Extract from this Atom XML: paper title, full abstract/summary, all author names (in order), the complete arXiv ID (strip only the URL prefix from <id>), the year from <published>, and the term attribute of <arxiv:primary_category>."

Retain legacy prefixes such as math/ and requested versions such as v2 in requests, metadata, eprint and URL.

For either lookup path, validate HTTP success and Atom structure. Report fetch, parse or Error entries (including their summary) as failures. Only a valid, successful empty feed establishes no result for the query.


Step 2b — Keyword search

WebFetch:

https://export.arxiv.org/api/query?search_query=<ENCODED-QUERY>&max_results=10&sortBy=relevance

URL-encode the entire all:<QUERY> parameter value, not just spaces: all:C++ becomes all%3AC%2B%2B. Check the feed's echoed query.

Prompt: "List the arXiv papers in this Atom XML: for each <entry> extract title, authors (all names), arXiv ID, and primary category. Return as a numbered list."

Present the numbered list to the user. Ask which paper(s) to retrieve in full (or proceed with all if ≤ 3 results). Then run Step 2a for each chosen ID.


Step 3 — Assemble BibTeX

From the extracted metadata, generate one @misc entry per paper:

@misc{<cite-key>,
  title        = {<title>},
  author       = {<Last, First and Last, First and ...>},
  year         = {<YYYY>},
  eprint       = {<ID>},
  archivePrefix = {arXiv},
  primaryClass  = {<primary-category>},
  url          = {https://arxiv.org/abs/<ID>}
}

Cite-key convention: <first-author-last-name><year><first-content-word-of-title>, all lowercase, no spaces. Example: dupuy2024basic.

Author format: invert each name to Last, First (or Last, First Middle); separate multiple authors with and.


Step 4 — Return

Output all four fields for each paper:

  1. Title
  2. Abstract (full text)
  3. Authors (comma-separated, natural order)
  4. BibTeX (fenced code block)

MCP enhancement (optional)

For full-paper download and reading, install blazickjp/arxiv-mcp-server:

uv tool install arxiv-mcp-server

Register in the project root's .mcp.json via:

claude mcp add --scope project --transport stdio arxiv-mcp-server -- arxiv-mcp-server

Restart; check claude mcp list for connection status and discover session tools.

When connected, the tools mcp__arxiv-mcp-server__search_papers, mcp__arxiv-mcp-server__download_paper, and mcp__arxiv-mcp-server__read_paper support download/read and local caching. Citation graphs use citation_graph; semantic_search additionally requires the optional arxiv-mcp-server[pro] install. The Atom path remains available.


Simple test

Query: arXiv ID 2407.19122
Expected title: "The Basic Theory of Clifford-Bianchi Groups for Hyperbolic n-Space"
Expected authors: non-empty list matching the arXiv API response
Expected BibTeX: @misc with eprint containing 2407.19122; retain requested version

04 同仓库更多 Skills