Scholar sources

日本語: 学術情報源

Property	Value
Plugin id	scholar-sources (bundled, off by default; the Research & academic template enables it for its Projects)
Access	Read-only toward every provider. No account, no API key, no credential
Writes	One: an open-access PDF into the Project’s corpus-inbox/, only through the attachment intake
Providers	arXiv API, OpenAlex, Crossref REST API, NCBI E-utilities (PubMed), doi.org content negotiation
Status	Available

Tools

Tool	What it does
scholar_search	query or up to 5 queries, sources (arxiv, openalex, crossref, pubmed; default all), since (YYYY-MM-DD), limit ≤ 25. Normalized records merged across sources by DOI, arXiv id, PMID and OpenAlex id
scholar_get	One paper by DOI, arXiv id, PMID or OpenAlex id: the full record with the abstract (up to 4,000 characters)
scholar_cite_check	Up to 50 citations, or the path of a Project draft (a file, or a folder: its newest documents). Each reference: verified, mismatch or not_found, with the reason. list: true lists draft versions; draft checks one version
scholar_fetch_pdf	The open-access PDF of one paper into corpus-inbox/ (see below). Returns the stored path, never the content
scholar_grant_deadlines	The researcher’s own deadline list (for example knowledge/grants.md) as reminders due today, 30, 14, 7, 3 and 1 days before each deadline and on the day; each reminder has a stable id

A record has: id (the DOI when there is one, else arxiv:, pmid: or openalex:), title, authors (at most 10, with the total), venue, date, every year the source gives, DOI / arXiv id / PMID / OpenAlex id, URL, open-access status, open-access PDF and licence when known, and an abstract snippet (at most 800 characters in a search). All provider text reaches the model inside the untrusted-content boundary; a boundary marker inside a title or an abstract is neutralised. Audit rows (connector.scholar-sources.<op>) carry ids, counts and byte sizes only, never a query, a title or an abstract.

The tools work in Project chats and in autopilot routines, which may call all five in gather (they are declared read tools of this plugin).

Providers and their etiquette

Verified against each provider’s documentation on 25 October 2026. Kurahi sends one request at a time per provider, with one limiter per provider for the whole gateway process, so parallel chats and routines share one budget.

Provider	Endpoint	What the provider asks	Kurahi
arXiv	export.arxiv.org/api/query (Atom)	No more than one request every three seconds, one connection, for all your machines together (terms)	1 request per 3 s; descriptive User-Agent
OpenAlex	api.openalex.org/works	10 requests per second without a key; keyless calls share one small daily budget per IP address (limits)	5 per second; no key
Crossref	api.crossref.org/works	Public pool 5 per interval, concurrency 1; polite pool (with mailto) 10, concurrency 3; back off on 429 (access)	3 per second, one at a time; mailto when configured
NCBI E-utilities	eutils.ncbi.nlm.nih.gov/entrez/eutils (esearch, efetch)	At most 3 requests per second without an API key; tool and email on every call (usage)	3 per second; tool and email on every call
doi.org	https://doi.org/<doi> with Accept: application/vnd.citationstyles.csl+json	Content negotiation for Crossref, DataCite, mEDRA and other agencies (docs)	3 per second; only for DOIs Crossref and OpenAlex lack

Every call retries at most twice on 429 and 5xx (honouring Retry-After up to 30 seconds) and then reports rate_limited or provider_unavailable. In a search, a source that refuses is named (Not available this time: arxiv (rate_limited)) and the other sources still answer.

Notes from the live recording: arXiv answered 429 Rate exceeded. (no Retry-After) to search queries although they were minutes apart, while id_list lookups succeeded; Kurahi then reports arXiv as not available and the other sources still answer. OpenAlex reports in every keyless response that the address shares a daily budget of US$0.10 (about 100 list requests). A weekly literature watch stays well inside both.

Configuration

{
  plugins: {
    entries: {
      "scholar-sources": {
        config: {
          contactEmail: "lab@example.ac.jp", // Crossref mailto, NCBI email, User-Agent
          ncbiTool: "kurahi-scholar-sources", // register tool + email with NCBI
          sources: ["arxiv", "openalex", "crossref", "pubmed"], // scholar_search default
          timeoutSeconds: 30,
        },
      },
    },
  },
}

Nothing is required: without contactEmail the tools still work (Crossref’s public pool, NCBI without email). With it, use the researcher’s or the lab’s address.

Citation check

scholar_cite_check resolves each reference by its identifier: a DOI through Crossref (/works/{doi}), then doi.org content negotiation (DataCite and other agencies); an arXiv id through the arXiv API; a PMID through PubMed. It then compares what the reference says with the record: at least 60% of the record’s title words must appear in the reference (when the reference has text besides the identifier), and a year written in the reference must be one of the record’s years (issued, online, print, first version). A reference without an identifier is searched in Crossref (query.bibliographic) and counts as verified only on a close match (80% of the title words and the year). A provider error is reported as “could not be checked”, never as a wrong reference.

From a draft, the tool reads the part after the last References heading (References, Bibliography, 参考文献, 引用文献, 文献), one entry per numbered or bulleted line.

Open-access PDFs only

scholar_fetch_pdf stores a PDF only when it is open access:

• an arXiv e-print (arXiv’s terms allow retrieval for research use; the copy stays in the researcher’s own Project and is not re-served), or
• a work whose best open-access location (OpenAlex) or Crossref licence is an open licence (Creative Commons or public domain) and names a direct PDF.

Anything else is refused before a byte is downloaded, including free-to-read copies without a licence. The download goes through the SSRF guard (public hosts only, at most 5 redirects), is capped at 20 MB and must start with %PDF- (a landing page is refused). The bytes then go through the attachment intake: quarantine, static policy, malware scan (fail closed), and only on a clean verdict a copy in corpus-inbox/ (a writable folder; 実行 Act mode). A refusal by the intake is audited and sealed like any attachment refusal, and in an autopilot run it flags the item malicious_attachment.

Grant deadlines

scholar_grant_deadlines reads one - YYYY-MM-DD | name | note line per deadline (lines in HTML comments are ignored) and computes the reminders due today in the given time zone (default Asia/Tokyo): a deadline n days away falls in the smallest distance at least n (30, 14, 7, 3, 1, or 0 on the day; remindDays changes the distances). Each reminder’s id (grant:<name>:<date>:d<distance>) is stable, so a routine that fans out over the ids files each reminder once. Nothing is submitted anywhere.

Never

Never submits a paper or an application, never contacts anyone, never downloads a paywalled copy, never sends a credential. See the Scholars never-list in Autopilot.