arxiv-search

name	arxiv-search
description	Search arXiv preprints for a topic, author, or concept and triage the results. Use when the user wants recent primary literature, preprints, or methods papers in physics, math, CS, quantitative biology, or statistics. Uses the plugin's built-in read-only arXiv MCP tool — no credential or install needed.
license	Apache-2.0

arXiv Literature Search (Codex-native)

Codex-native literature search wired to this plugin's built-in read-only MCP tool. Inspired by Google DeepMind's literature-search-arxiv (science-skills, Apache-2.0); the MCP integration is Codex Science's own.

How to run

1. Search with the plugin MCP tool science_search_arxiv:
   • arguments: query (topic/author/keywords, ≤500 chars), limit (1–10, default 5).
   • returns a JSON list of {id, title, summary, url} — arXiv id, title, abstract, and the abstract-page URL.
2. Triage the returned abstracts against the user's actual question. Discard off-topic hits; do not pad the list.
3. Refine if needed: re-query with more specific terms, an author, or a narrower concept. arXiv full-text search is broad, so precise terms help.
4. Go deeper only when justified: to read a paper, open its url (or the /pdf/ variant) with a fetch tool and ask before large downloads.
5. Report each kept paper with title, arXiv id, one-line relevance, and the URL. Separate strong matches from tangential ones.
6. Provenance & review: record the queries, tool, and cited ids with $science-provenance; check claims with $science-review before presenting.

Boundaries

• arXiv preprints are not peer-reviewed; flag that when it matters and prefer the published version where one exists.
• Cite only ids/titles actually returned by the tool — never invent arXiv ids or fabricate abstracts.
• The tool is discovery, not full text; base specific claims on the paper you actually read, not the abstract alone.