datacite-mcp-server

---

Overview

DOI metadata from DataCite for datasets, software, samples, workflows, and other research outputs that repositories deposit worldwide. Search it, open a record in full, trace a DOI's relations (versions, parts, supplements, citations), find the repositories that publish in a field, and format citations. No API key required.

Tools

Tool	Description
datacite_search_works	Search works by plain text or query syntax plus structured filters, in ranked pages or a full cursor walk, with optional facet counts
datacite_get_work	Fetch the full deposited metadata for one DOI, or learn which registration agency holds a DOI DataCite doesn't
datacite_trace_relations	Map the relation graph around any DOI — versions, parts, supplements, derivations, citations — with each edge's source
datacite_search_repositories	Find repository accounts by text, field of science, type, certificate, software, or provider, and get the repositoryId work search filters on
datacite_get_citation	Render a DOI as a formatted citation in a CSL style and locale, or as BibTeX, RIS, CSL JSON, and other machine formats
datacite_list_reference	Look up the vocabularies, identifier forms, and coverage rules the other tools accept, offline

Capability reference

datacite_search_works _tool

• text (literal words, with angle brackets removed; nonblank text that leaves no searchable clause fails as invalid_text) or query (OpenSearch query syntax), ANDed with filters: resource_types, creator (ORCID iD or name), affiliation (ROR ID or name), affiliation_country, funder (ROR ID, Crossref Funder ID, or name; include_child_funders with a ROR funder), subject, fields_of_science, repository_ids, provider_ids, licenses, language, place, published_from / published_to, and min_citations; list filters take up to 10 values, any of which match
• limit 1–10,000 (default 20), applied at most 100 with a notice; ranked pages reach the first 10,000 matches under sort (relevance, newest, oldest, recently_updated, most_cited, most_viewed, most_downloaded; default relevance with text or query, else newest), while cursor: "*" walks the whole result set in registration order through nextCursor and takes neither page nor sort. A ranked ceiling stop is distinguished from the last page
• Crossref Funder IDs resolve through one exact ROR FundRef mapping before searching funded-by; an unavailable, incomplete, or ambiguous mapping keeps literal identifier matching and discloses reduced coverage. The cached, paced lookup takes at most five seconds without retry. A cursor walk keeps its initial resolved or literal filter on every continuation without another ROR lookup, including across server restarts. include_child_funders still requires an explicitly supplied ROR
• License ids are trimmed and lowercased, preserving internal spaces (cc0 1.0 differs from cc0-1.0); reuse the licenses facet ids exactly. query also supports deposition timestamps, e.g. created:[2026-10-16 TO *], updated, and registered
• include_facets adds top counts for resource types, years, repositories, providers, affiliations, fields of science, and licenses (2–8 s slower); every response reports totalCount, effectiveQuery, sortApplied, and appliedFilters

---

datacite_get_work _tool

• One doi, in any case: bare, with a doi: or info:doi/ prefix, as a doi.org URL, or %2F-encoded
• Titles, creators and contributors with ORCID iDs and ROR affiliations, dates, subjects, descriptions, funding, geolocations, rights, related identifiers and items, repository and provider, and counts (citations, references, versions, parts, views, downloads); long lists are capped and their full sizes reported in truncatedLists and relatedIdentifierCounts
• Page one capped list with a zero-based offset (default 0): creators, contributors, relatedIdentifiers, subjects, and fundingReferences return up to 100 entries, relatedItems 25, and geoLocations 50. Offsets index valid mapped entries in deposit order. listPage reports list, offset, total, and nextOffset while entries remain; other metadata keeps its normal caps. Default truncation notices give exact continuation inputs. An offset requires a list; past-end pages are empty
• Titles and search snippets are plain text, with recognized formatting tags removed and entities decoded once before snippet truncation. Structured descriptions keep their deposited text; rendered descriptions preserve paragraph and line boundaries
• Geolocations preserve points, boxes, ordered polygon vertices, and optional polygon interior points. Incomplete polygons are omitted with a notice; valid sibling geometry remains
• A DOI DataCite holds no public record for returns found: false with missReason (other_agency, does_not_exist, not_public, unclassified), the registrationAgency when known, and guidance

---

datacite_trace_relations _tool

• Any agency's doi: a DataCite dataset or software DOI, or a journal article's DOI to find the DataCite data and software it cites or that cite, supplement, or derive from it
• depth 1 (default) or 2 — the second hop expands at most 10 DataCite neighbours and runs only when the first hop leaves max_nodes room; max_nodes 1–10,000 (default 50, root included), applied at most 100 with a notice, fills with own-metadata targets, then records pointing at the root, then Event Data endpoints, then the second hop; relation_types keeps only the listed relation types, read from the traced DOI's side (omitted: all); include_event_data (default true) adds Event Data citation links
• Edges keep relationType exactly as asserted and list their sources (metadata, reverse_metadata, event_data); coverage states how much of each source was read (up to 100 reverse records per hop, only the first 10 when those are large, and up to 100 events per call), and rootCounts carries DataCite's own counts for comparison. An absent edge is not evidence that no relationship exists
• Reverse lookups inspect both relatedIdentifiers and identified relatedItems, verify each assertion's target and type, and merge identical edges
• event_offset (0–9,999, absolute and zero-based) selects Event-Data-only tracing, including when set to 0: metadata edges and depth-2 expansion are skipped. Continue with coverage.eventData.nextEventOffset and the same DOI and filters. Each call reads one upstream page and stops before an eligible endpoint that cannot fit max_nodes; root-only calls explain how to raise the cap. Coverage distinguishes exhaustion, a blocked classification lookup, unavailable/skipped reads, and the upstream 10,000-result ceiling. Upstream ordering can change between calls; continuations are not snapshots

---

datacite_search_repositories _tool

• query text over names and descriptions, plus field_of_science, repository_types, certificates, software, client_type, and provider_id; or repository_ids alone (up to 1,000 inputs) to look up the first 25 distinct normalized ids in input order. A notice names unqueried ids; missing-id notices cover queried ids only
• limit 1–10,000 (default 20), applied at most 100 with a notice, with page; name-ordered rows carry the repositoryId that datacite_search_works takes in repository_ids, plus providerId, types, certificates, software, subjects, homepage, and re3data link
• Search responses echo the composed effectiveQuery (* without query clauses), including empty pages; ID lookups report appliedFilters.repository_ids without a query echo

---

datacite_get_citation _tool

• format: text (default), csl_json, bibtex, ris, datacite_json, datacite_xml, schema_org, codemeta, or jats
• For text only: style takes a CSL style id (default apa; verified ids under datacite_list_reference topic citation_styles), and one DataCite would silently render as APA fails as unsupported_style; locale takes one of the 61 CSL locales DataCite renders (en-GB, de-DE, fr-FR, …) or a bare language code that expands to its primary dialect (de → de-DE), default en-US (list: topic citation_locales)
• text returns the citation as plain text in citation and the upstream markup in citationHtml; text payloads over 100,000 UTF-16 code units are cut without splitting a surrogate pair. Machine formats return complete payloads inline up to that cap; larger documents omit citation and return downloadUrl for the full document and totalCharacters, with truncated: true, shown: 0, and cap: 100000. A DOI DataCite doesn't hold returns found: false as in datacite_get_work; a format DataCite can't render for that DOI fails as format_unavailable

---

datacite_list_reference _tool

• topic: resource_types, relation_types, identifier_types, date_types, contributor_types, fields_of_science, licenses, repository_types, certificates, software_platforms, client_types, sort_orders, query_syntax, citation_formats, citation_styles, citation_locales, identifier_formats, or coverage
• Returns entries (value, label, group, inverse) and usage notes; makes no upstream request

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

The published datacite-guide skill covers attribution, identifier resolution, relation evidence, and complete inventories. It ships with the package and plugins and is available through the MCP Skills extension.

DataCite-specific:

• Keyless and read-only: the DataCite REST API (/dois, /repositories, Event Data at /events, and DOI content negotiation for citations), plus the doi.org registration-agency lookup for DOIs DataCite doesn't hold
• Shared request budget: one paced request queue, an in-process response cache, and a 45 s deadline per tool call; a spent budget fails as rate_limited with the wait in seconds. DATACITE_CONTACT_EMAIL moves requests to DataCite's identified tier, 1,000 requests per 5 minutes per IP instead of 500
• Forgiving identifiers: DOIs in any case, bare or prefixed or as doi.org URLs; ORCID iDs with their checksum verified; ROR IDs and Crossref Funder IDs bare or as URLs; vocabulary values in any case, with hyphens, underscores, or spaces
• DataCite DOI metadata is waived under CC0 1.0; the waiver covers the deposited metadata only, not the datasets, software, or papers it describes. datacite_get_work reports metadataLicense separately from the work's own rights

Agent-friendly output:

• Provenance on every search: effectiveQuery, appliedFilters as sent upstream, the sort or order applied, and totalCount; trace edges name their sources, and coverage says how much of each source was read
• Typed misses instead of errors: a DOI DataCite doesn't hold comes back found: false with a missReason, the agency that holds it when known, and the next step
• Disclosed limits: capped lists report their full counts, and notices explain zero hits, the 10,000-match paging ceiling, an unexpanded second hop, and citations the returned edges don't account for
• In the text output, depositor-supplied text (titles, descriptions, names) is blockquoted or flattened and machine payloads are fenced, so deposited content can't pose as structure. Titles and snippets are normalized on both surfaces; structured descriptions and other deposited fields retain their original text

Getting started

Connect to the hosted instance, no install:

claude mcp add --transport http datacite-mcp-server https://datacite.caseyjhand.com/mcp

Every caller of the hosted instance shares one DataCite request budget of 800 requests per 5 minutes. For sustained use, run your own instance.

Or run it locally:

claude mcp add datacite-mcp-server -- npx -y @cyanheads/datacite-mcp-server

For other clients, add this to the MCP config file:

{
  "mcpServers": {
    "datacite-mcp-server": {
      "command": "npx",
      "args": ["-y", "@cyanheads/datacite-mcp-server"]
    }
  }
}

No API key is needed. Set the optional DATACITE_CONTACT_EMAIL to double DataCite's per-IP request allowance. For Docker or a source build, see Running the server.

Configuration

Variable	Description	Default
DATACITE_CONTACT_EMAIL	Contact email sent in the User-Agent as mailto:, never in a URL. Moves requests to DataCite's identified tier: 1,000 requests per 5 minutes per IP instead of 500. Validated as an email at startup.	none
DATACITE_MAX_REQUESTS_PER_5MIN	Request budget per 5-minute window, an integer from 50 to 1000. Divide it across replicas that share one egress IP.	800 with a contact email, 400 without
MCP_TRANSPORT_TYPE	Transport: stdio or http.	stdio
MCP_HTTP_PORT	HTTP server port.	3010
MCP_SESSION_MODE	HTTP session mode: stateless, stateful, or auto.	stateless
MCP_AUTH_MODE	Authentication: none, jwt, or oauth.	none
MCP_LOG_LEVEL	Log level (debug, info, warning, error, etc.).	info
LOGS_DIR	Directory for log files (Node.js only).	<app-root>/logs
STORAGE_PROVIDER_TYPE	Storage backend: in-memory, filesystem, supabase, cloudflare-kv/r2/d1.	in-memory
OTEL_ENABLED	Enable OpenTelemetry.	false

See .env.example for the full list of optional overrides.

Running the server

Requires Bun v1.4.0 or higher, or Node.js v24+. To build from source:

git clone https://github.com/cyanheads/datacite-mcp-server.git
cd datacite-mcp-server
bun install
cp .env.example .env

Set DATACITE_CONTACT_EMAIL in .env to use DataCite's identified tier.

Local development

• Build and run the production version:

  # One-time build
  bun run rebuild
  
  # Run the built server
  bun run start:http
  # or
  bun run start:stdio
  

• Run checks and tests:

  bun run devcheck  # Lints, formats, type-checks, and more
  bun run test      # Runs the test suite
  

Docker

Run over stdio:

docker run -i --rm -e MCP_TRANSPORT_TYPE=stdio ghcr.io/cyanheads/datacite-mcp-server:latest

Or expose Streamable HTTP at http://localhost:3010/mcp:

docker run --rm -p 127.0.0.1:3010:3010 ghcr.io/cyanheads/datacite-mcp-server:latest

Project structure

Directory	Purpose
src/index.ts	createApp() entry point — registers the six tools and sets the server instructions.
skills/	Published guidance for agents using the DataCite tools.
src/mcp-server/tools	Tool definitions (*.tool.ts) and the input-schema, miss-guidance, and text-rendering helpers they share.
src/services/datacite	DataCite REST client (/dois, /repositories, /events, content negotiation), query builder, identifier normalizers, record mappers, relation-graph builder, and citation style check.
src/services/doi-ra	doi.org registration-agency lookup for DOIs DataCite holds no public record for.
src/services/ror	Exact Crossref Funder ID to ROR lookup with literal-search fallback.
src/services/http	Shared upstream pipeline — response cache, request pacer, retry, per-call deadline, rate-limit handling.
src/services/reference	Static vocabularies behind datacite_list_reference and the input validators.
src/config	Server-specific environment variable parsing and validation with Zod.
tests/	Vitest tests mirroring src/, with recorded DataCite and doi.org fixtures.

Development guide

See CLAUDE.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic
• Use ctx.log for logging, ctx.state for storage
• Register new tools in the createApp() arrays in src/index.ts
• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheck
bun run test

License

This project is licensed under the Apache 2.0 License. See the LICENSE file for details.