@pipeworx/research-chain

Connect: https://gateway.pipeworx.io/research-chain/mcp · Install: one-click buttons

Tools: 1

Everything openly available around one academic paper, in a single call: full text, license, datasets, code, and citations. Given a DOI, PMID, or arXiv id, research_chain walks the join that Crossref, OpenAlex, Unpaywall, DataCite, Zenodo, Hugging Face Papers, PubMed Central, and Semantic Scholar each expose piecemeal — so a caller no longer has to know all six/seven tools to answer “what datasets and code are linked to this paper?” or “find an open copy of this DOI and its license.”

Tools

• research_chain({doi | pmid | arxiv_id, max_datasets?, max_citing?}) — returns:
  • paper — title, journal, publisher, authors, Crossref’s stated license(s) (URL + which manuscript version each applies to), OpenAlex concepts and citation count.
  • open_copies[] — every open-access location Unpaywall knows about, each with its OWN host_type, version (publishedVersion | acceptedVersion | submittedVersion), and license — never collapsed. is_oa: true only means “a free-to-read copy exists somewhere”; it says nothing about reuse rights, which is why license is null (not guessed) whenever Unpaywall doesn’t state one. A PubMed Central copy (if one exists) is added as its own entry with license and version explicitly null, since NCBI’s ID Converter confirms the copy exists but doesn’t state its terms.
  • pmc_full_text — {has_full_text, pmcid, pmid, url} from NCBI’s ID Converter.
  • datasets — {total, returned, results[]} from DataCite’s relatedIdentifiers search (datasets/software that cite the paper’s DOI). total is the real upstream count even when results is truncated to max_datasets.
  • code — {zenodo[], paperswithcode}. Zenodo is searched by metadata.related_identifiers.identifier. paperswithcode (Hugging Face Papers, the de-facto successor) reports {found: false, results: []} when the upstream search comes back empty — an empty result is not_found, never treated as an answer (a bare [] from this upstream used to be routed to as if it were a real “no code exists” answer).
  • citing — {count, top[], note} from Semantic Scholar: count is the paper’s total citation count, top is up to max_citing citing papers sorted by their own citation count. note carries an honest partial- failure message (e.g. a rate limit) when one of the two Semantic Scholar calls failed but the other succeeded — it is never silently dropped.
  • retraction — derived from the SAME Crossref record already fetched for paper (title marker, update-to retraction notices, is-retracted-by relation) — no extra API call. retracted: false means “no retraction on record”, not proof the paper is clean (coverage depends on the publisher).
  • sources_used / sources_failed — every leg fetches independently and in parallel; a slow or dead upstream lands in sources_failed with a reason, and every other leg still returns.
  • as_of — ISO timestamp of the call.

Auth

Keyless. Every upstream here is a public, unauthenticated API. Crossref, OpenAlex, and the NCBI ID Converter ask only for a polite-pool contact email, which this pack supplies as its own ([email protected]) — the same pattern as the crossref, openalex, and unpaywall packs.

Data sources

• https://api.crossref.org/works/{doi} — title, journal, license, retraction markers.
• https://api.openalex.org/works/doi:{doi} — concepts, citation count.
• https://api.unpaywall.org/v2/{doi} — open-access locations, each with version + license.
• https://pmc.ncbi.nlm.nih.gov/tools/idconv/api/v1/articles/ — DOI/PMID ↔ PMCID resolution. NCBI moved this off www.ncbi.nlm.nih.gov/pmc/utils/idconv in 2026 — the old host 301-redirects here; this pack hits the new host directly.
• https://api.datacite.org/dois — datasets/software citing the paper (relatedIdentifiers.relatedIdentifier search).
• https://zenodo.org/api/records — records related to the paper. The search field is metadata.related_identifiers.identifier, not the bare related_identifiers.identifier — the bare field silently matches zero records for every query (verified live: 0 vs 19 hits for the same DOI).
• https://huggingface.co/api/papers/search — code/implementation search (successor to the shut-down Papers with Code API).
• https://api.semanticscholar.org/graph/v1/paper/DOI:{doi} and /citations — citation count + citing papers. Keyless pool is ~1 req/sec cumulative across BOTH endpoints; this pack sequences them with a ~1.1s stagger inside one leg rather than firing both in the same instant, which 429’d one of them on nearly every call before the fix.

OSF was scoped as a third code/materials source but dropped: its public API has no DOI-filterable field (/v2/preprints/?filter[doi]=... returns HTTP 400 “not a filterable field”, verified live). Zenodo + Hugging Face Papers cover the code/materials join for now.

Identifier resolution

• doi — used directly (a doi.org URL or doi: prefix is stripped).
• pmid — resolved to a DOI via NCBI’s ID Converter.
• arxiv_id — resolved to arXiv’s own auto-registered DataCite DOI (10.48550/arXiv.<id>), which covers essentially every arXiv submission since 2022. This is a construction, not a verified lookup — a pre-2022 arXiv id without a registered DOI will report not_found on the downstream legs rather than resolving.

Tools

• research_chain — Everything openly available around a paper, in ONE call: full text, license, datasets, code, and citations. Given a DOI, PMID, or arXiv id, walks paper (Crossref + OpenAlex) -> open copies with PER-CO