Find research datasets with an AI agent and DataCite MCP

Run the workflow

1. 1. Define the dataset question and search phrase

   Describe the topic, geography, observation period and variables you need. Choose a short phrase likely to appear in deposited metadata, such as climate finance. datacite_search_datasets treats the phrase literally: Boolean and field operators are not exposed. Try terminology variants in separate searches and record each input. The optional fromYear and toYear bounds apply to the publication year, not the period covered by the underlying observations.

2. 2. Search and preserve coverage limits

   Call datacite_search_datasets with query, limit, page and optional inclusive publication-year filters. The default limit is five and the maximum is ten. Keep the query, filters and limit unchanged when using nextPage; pages stop at 100. Preserve paginationLimitReached and the search time. DataCite search only covers public Findable DOI records classified as Dataset, so it is not a search across every repository or a complete inventory for your topic.

3. 3. Inspect each candidate DOI and its deposited rights

   Use datacite_dataset_details with a plain DOI, not a doi.org URL. Compare titles, creator names, publisher, publication year, subjects, rights and related identifiers. Keep missingMetadata, list totals, truncated flags and textTruncated in your notes. This connector returns no descriptions or dataset files. Missing rights must remain unresolved; a deposited rights identifier or URL is a claim to check at the repository, not proof of unrestricted reuse.

4. 4. Investigate an institution only when relevant

   If a repository or another source supplies an institution name, investigate it separately with ror_search_organizations. Compare candidate names, country, organization type and websites before selecting a record; do not automatically choose the first result. Inspect a selected plain nine-character ID with ror_organization_details and retain its status. DataCite's ToolCargo output contains creator names but no creator affiliations, so these tools do not establish a dataset-to-institution relationship. ROR inclusion does not establish accreditation or dataset quality.

5. 5. Review the actual repository before selecting data

   Follow the DOI source separately and check the dataset version, documentation, variables, coverage, access restrictions and applicable license. Review related identifiers in context: a relation can describe another version, a publication or a different research object. Record what you actually inspected and which requirements remain unknown. The hosted tools do not download the data, run analyses or judge fitness for your research question.

What a useful result looks like

A useful shortlist has one entry per dataset DOI: source link; title; creator names; publisher; publication year; deposited rights and related identifiers; missing or truncated metadata; and the repository checks still needed. Keep publication dates separate from observation dates. Add an institution identity only with a separately recorded source for the relationship and the inspected ROR record. A complete shortlist can contain unresolved access or license questions.

Know the limits

Metadata discovery only: public Findable DataCite Dataset DOIs, literal phrase searches, at most ten records per page and 100 pages. Records, counts and order can change. No descriptions, full text, dataset downloads, automatic affiliation matching or quality assessment. ROR is an optional separate name search; its default projection is five of up to 20 records per upstream page. Use limit 20 to avoid local omissions before moving to nextPage. Check missing fields and truncation in both connectors. Hosted calls require a ToolCargo account, authentication and available shared quota; no DataCite or ROR provider key is needed.