# FAIRdata.ai

> Research data, prepared for the machines that use it. FAIRdata.ai turns public
> academic datasets into evidence-checked, machine-readable records, collections
> and ingestion recipes, so models, agents and research workflows can find,
> understand and use them more responsibly.

## What this is

A public registry of research datasets, assessed for FAIR metadata quality and
curation evidence, with normalised licence signals and resolved file URLs. The
qualifying subset is organised into themed collections with versioned manifests.

Public read access requires no authentication. Rate limit: 240 requests per 60
seconds per client IP, reported in RateLimit-* headers.

## What "qualified" means here

A dataset is Machine-Ready when: source-record FAIR composite >= 75 AND curation-evidence proxy >= 50 AND an unambiguous permissive licence signal AND at least one declared file format AND a direct file URL resolved at assessment time.

Read the full ruleset — every criterion, threshold, evidence type and status —
at https://fairdata.ai/api/v1/qualification/ruleset.

## What it does NOT mean

Machine-readable does not mean permission to train. Qualification does not
establish that data are scientifically correct, representative, free of personal
or sensitive information, covered by appropriate consent, or legally available
for any particular use. Licence detection and discipline classification are
automated and can be wrong. File access is resolved point-in-time.

Every qualification response includes a "limits" array. If you relay a status,
relay the limits with it.

## Endpoints for agents (no key required)

- GET https://fairdata.ai/api/v1/meta — capabilities, collection version, ruleset, formats, error catalogue
- GET https://fairdata.ai/api/v1/health — service and data freshness
- GET https://fairdata.ai/api/v1/collections — collections with real aggregates
- GET https://fairdata.ai/api/v1/collections/{id} — one collection
- GET https://fairdata.ai/api/v1/collections/{id}/qualification — per-criterion evidence
- GET https://fairdata.ai/api/v1/collections/{id}/datasets — members, filtered and paginated
- GET https://fairdata.ai/api/v1/collections/{id}/manifest — versioned manifest with checksum
- GET https://fairdata.ai/api/v1/collections/{id}/croissant — Croissant 1.0 JSON-LD
- GET https://fairdata.ai/api/v1/collections/{id}/receipt — fairdata.lock.json
- GET https://fairdata.ai/api/v1/collections/versions — published versions and retention
- GET https://fairdata.ai/api/v1/datasets/resolve?doi={doi} — resolve a DOI
- GET https://fairdata.ai/api/v1/search?q={query} — substring search over qualifying datasets
- GET https://fairdata.ai/api/record/{doi} — the full enriched record (whole registry)
- GET https://fairdata.ai/api/record/{doi}/{schema.org|croissant|ro-crate|fdp|datacite|ai-readiness}
- GET https://fairdata.ai/api/record/{doi}/files — resolved file listing
- GET https://fairdata.ai/api/search/semantic?q={query} — meaning-based search over the whole registry

Errors are RFC 9457 problem+json with a stable "code". A JSON endpoint never
returns HTML. 503 snapshot_unavailable means the data could not be read — it is
not an empty result, and must not be reported as "no collections".

## MCP

Endpoint: POST https://fairdata.ai/api/mcp (Streamable HTTP, JSON-RPC 2.0)
Discovery: https://fairdata.ai/.well-known/mcp.json
Protocol versions: 2026-07-28 (current), 2025-11-25, 2025-06-18, 2025-03-26, 2024-11-05

Tools: get_ruleset, search_collections, get_collection, compare_collections,
list_collection_datasets, get_collection_manifest, explain_collection_qualification,
search_datasets_in_corpus, get_dataset, explain_dataset_qualification,
create_ingestion_plan, create_reproducibility_receipt, check_usage_constraints,
plus the record tools get_record, get_record_format, list_records,
get_ai_readiness, get_files, get_data_quality, find_similar_datasets,
search_datasets, get_linked_papers, assess_dataset.

check_usage_constraints returns evidence and unknowns, never a yes/no on whether
a use is permitted.

## Documentation

- https://fairdata.ai/developers — the developer portal
- https://fairdata.ai/developers/index.md — this documentation as Markdown
- https://fairdata.ai/api/openapi.json — OpenAPI 3.1
- https://fairdata.ai/schemas/collection-manifest-1.0.0.json
- https://fairdata.ai/schemas/fairdata-lock-1.0.0.json

## How to cite

Cite the dataset, not FAIRdata.ai — we describe and assess these records, we did
not create the data. If you want to record how a selection was made, cite the
collection id, its version and the manifest checksum.

## Part of

Infinite Researchers (https://infiniteresearchers.com). Sister projects:
OpenScience.ai, Preprints.ai, OpenAccess.ai.

Contact: hello@fairdata.ai