papers external

Live arXiv paper search via papers-mcp (github.com/guangxiangdebizi/papers-mcp), a stdio MCP server run as a local subprocess. Despite the project's own multi-source branding (OpenAlex+Crossref+arXiv+Unpaywall), the live papers.search tool queries arXiv only as of 2026-07-29 verification -- the OpenAlex/Crossref code paths exist in the server's source but are never called.

tier: external
category: scholarly
coverage: live
ranking: none -- papers.search returns no relevance/rank number per hit. List order reflects arXiv's own (recency-leaning) result ordering, but that ordering is not exposed as a comparable score field.
knobs: q, limit, authors, year_from, year_to, fields_of_study

Guidance

arXiv preprints only, live from arXiv's own API -- this is the differentiator vs the `openalex` corpus, which is a frozen 2026-07 snapshot of ~503M records across all venues but supports point-in-time `before` filtering that this corpus has no equivalent for. Hits are metadata + abstract only (`text` is abstract_snippet, not full text); `meta.oa_pdf_url` is a direct PDF link (arXiv is inherently open access) vs openalex's `landing_url`. `date` is a bare publication YEAR (e.g. 2023), not a full ISO date -- arXiv's Atom feed carries no month/day granularity here. doi is never populated on these hits (arXiv preprints as surfaced by this tool carry no DOI). year_from/year_to and fields_of_study (arXiv category codes, e.g. cs.LG) are real filters verified live to change result sets; note the tool's own `total` count is computed BEFORE year_from/year_to is applied (that filter runs client-side inside papers-mcp), so total can overstate the filtered set. authors does affect results (appended as an arXiv au: term) but is a fuzzy OR-ish match, not a strict author filter. Verified live 2026-07-29: affiliations, venue, open_access_only, and sort_by are accepted by the tool's own schema but have ZERO effect on results in the current shipped code (dead parameters -- searchArxiv() never reads them) -- do not rely on them; they are deliberately left off `native` here rather than advertised as knobs that silently do nothing. Keep offset+limit <= 100: papers-mcp clamps its own page_size to 100 server-side, so larger windows silently return fewer (or zero) hits rather than erroring.

Usage

The public endpoint is not live yet. Keys are invite-only during beta; contact shapeandfact+takedown@gmail.com.

Plain HTTP:

curl -s https://bipsearch.ai/search \
  -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{"corpus": "papers", "q": "transformer attention mechanism", "limit": 5}'

Or via MCP, after adding the server once (see quickstart):

# after: claude mcp add --transport http bipsearch https://bipsearch.ai/mcp \
#   --header "Authorization: Bearer YOUR_KEY"
list_corpora()
search(corpus="papers", q="transformer attention mechanism", limit=5)