mdexa-ai/mdexa-ai-native-database
DataDated, evidence-backed license and provenance cards for medical AI datasets, queryable by agents over MCP and HTTP.
What it is
Mdexa is an open snapshot of verified license and provenance records for the datasets and open models behind medical AI. It addresses a concrete problem: hosting pages, repositories, papers and terms pages often disagree about the same dataset, and a repository's code license can easily be mistaken for a data license.
Each resource becomes one dated record, with every license statement quoted verbatim and tied to a snapshot listed in evidence-index/ by address, retrieval time, SHA-256 and size. The repository ships cards, schema, decision rules and a validator; it contains no dataset files and does not publish the third-party snapshot text itself.
Coverage includes modern medicine, traditional Chinese medicine and cross-cutting areas. Records are bilingual.
How records are decided
Two status fields are kept separate and never collapsed:
license_statusstates what the sources say:permissive,restricted,conflict,unstatedorunverifiable.hostingstates what Mdexa does with the dataset:link-only,pending,eligibleorhosted.
risk_flags capture issues a license line cannot, such as exam-derived questions, scraped web content, real patient text or model-generated text. A permissive result plus a risk flag does not mean the dataset is safe to redistribute, and no dataset is currently hosted. Repository code licenses are recorded separately in repo_license. The decision process is published in RULES.md and as machine-readable rules in the schema, with a field reference in English and Chinese.
Quickstart
For batch analysis, read the export file exports/cards.jsonl, one card per line. Python example from the README:
import json
with open("exports/cards.jsonl", encoding="utf-8") as f:
cards = [json.loads(line) for line in f]
for c in cards:
if c["license_status"] == "permissive" and c["risk_flags"]:
print(c["slug"], c["risk_flags"], c["last_verified"])
The JavaScript variant requires Node 18 or later. To fetch a single live record over HTTP:
curl "https://mcp.mdexa.com/v1/datasets/cmb"
The matching evidence endpoint is /v1/datasets/cmb/evidence, and combined filters are supported in list queries. When calling from Python's urllib, set your own User-Agent; the default agent is refused. For agent use, point an MCP client at https://mcp.mdexa.com/mcp:
{
"mcpServers": {
"mdexa": {
"url": "https://mcp.mdexa.com/mcp"
}
}
}
Runnable Python, TypeScript and curl examples live in the separate MCP examples repo. To check the cards locally, run node tools/validate-cards.mjs, which needs Node 22 and no dependencies. It validates every card against the JSON Schema, the controlled taxonomy and the rules.
Limits and licensing
- The cards record what sources state; they are not legal advice, and a permissive verdict is not a clear bill for redistribution when risk flags are present.
- Light HTTP use needs no account, but the live database's search, change alerts and accounts are on the website; this repository is exported at each release, so counts here lag live statistics.
- Snapshot text is withheld because it is third-party content; only hashes and metadata are published.
- Cards and aggregate data are CC BY 4.0, while scripts and schema are MIT. Quotations and dataset names remain with their rights holders; see NOTICE.md for scope and corrections.
HotRepo's take
The repository first appeared on HotRepo's new-project list on 2026-11-17, gaining 20 stars in a roughly 1.5-hour observation window and reaching a total of 127 stars with no forks in the recorded snapshot. That is a short window and cannot support a trend judgment. The design choice worth noting is structural: separating what a license allows from what the hosting service does, and publishing per-conflict quotations plus a validator, makes the dataset usable as a lookup for agents as well as by human reviewers.
Written by AI from the project's README, updated . The project's own docs are the source of truth.
You might also like
- chinese-poetry/chinese-poetryA JSON dataset of classical Chinese poetry, covering Tang and Song poems and Song ci for app developers and researchers.
- lexifina/global-lawBrowse many countries' laws as one dated git history, querying what text was in force on any day.
- yihui-dev/how-profitable-startups-growA weekly-updated table of 742 startups with revenue bands, growth trends, founders and customer acquisition notes.