Get started Console

Modern medicine75

Traditional medicine24

Cross-cutting12

Collections

Models8

License status and Mdexa hosting, recorded separatelyAll datasets

Data infrastructure for medical AI

Every dataset has its license checked, each verdict comes with the source text, and people and agents can call it directly.

Explore the databaseSee a card
By verification level
96Verified5Practical card3Registered, not verified8Models
Evidence · excerpts checked
426404 with excerpt401 match snapshot
Source snapshots
1,098huggingface 602github 233
Cards updated, 7 days
101
Systems · domains
3 · 16

Last verified offline values

One question, two ways to answer it

“Chinese medical exam benchmarks licensed for commercial use”

Without Mdexathe agent opens every page
  1. huggingface.co/…/medical-exam-question-…Apache-2.0
  2. huggingface.co/…/LingLanCC-BY-4.0
  3. huggingface.co/…/include-base-44Apache-2.0
  4. huggingface.co/…/TCMLEApache-2.0
  5. huggingface.co/…/tmmluplusMIT

20 more pages

25
pages
5
sites
8
kinds of license location
  • 21 datasets, each read source by source
  • 31 license passages to read and compare
With Mdexaone call
search_datasets({ content_type: "benchmark-exam", language: "zh" })

16 more cards

1
call
21
cards
41
evidence entries
  • 21 results carry license_status and hosting
  • 21 carry last_verified and card_url

Under the hood

From source page to agent. Every figure below is counted from build artifacts.

https://huggingface.co/datasets/TCMLLM/LingLan

  1. Every verdict traces to a verbatim excerpt, an archived snapshot and a date.

    401 / 402 excerpts matched in their snapshot by script

  2. License facts and Mdexa hosting are two fields; the rules are public.

    5 × 4 license_status × hosting · 26 rules

  3. Same input, byte-identical output.

    two publish runs, one sha256 abf653e9…

  4. Built for agents and tested with real clients.

    700 chars per compact row · calls on a missing dataset: 8, now 1 · SDK 1.32.2 · Inspector 2.13.0 · 9 questions, 0 severe errors

  5. Machine-readable throughout.

    4 × 101 card files (JSON, Markdown × 2 languages) · catalog.json · llms.txt · JSON-LD

  6. Keys and tokens are stored only as hashes; call logs hold no address in plain text.

    SHA-256

About Mdexa

  1. Database

    Biomedical datasets, each with a license verdict and the source text behind it.

    LiveOpen
  2. Models

    Open biomedical models. Code license and weights license are checked separately.

    LiveOpen
  3. Physician-reviewed data

    Clinical judgments written and cross-checked by licensed physicians.

    In preparationHow
  4. Custom data

    Datasets built to specification, with the same license and review records.

    On requestContact

Anatomy of a card

A real card: ClinicalTrials.gov, verified 2026-11-13.

Clinical medicine

ClinicalTrials.gov

Commercial use
UnclearUnstated
Derivatives
Conditional
Redistribution
UnclearUnstated
File on Mdexa
NoPending
1
The verdict

Four questions a team asks first, answered from the evidence. Shape, not colour, carries the answer.

  1. Terms of use pagePage opened
    Third-party text · Terms of use pageThe following Terms and Conditions apply for use of ClinicalTrials.gov data. Downloading ClinicalTrials.gov data indicates your acceptance of the following Terms and Conditions.

    Statement found in snapshot s09 (terms-page)

    clinicaltrials.gov2026-11-13 · s09
2
The source text

Quoted verbatim, with the source URL, the date checked and snapshot s09. 10 pieces of evidence in all.

Known defects and misuse

  • Tse et al.: many records are not updated in time; as of February 2018 about 9.4% of records had 'unknown' recruitment status, meaning not updated or verified for two years.
    pmc.ncbi.nlm.nih.gov ·
    Despite the capability to make changes to a record, many records are not updated in a timely way.

    https://pmc.ncbi.nlm.nih.gov/articles/PMC5968400/ · peer-reviewed-paper · checked 2026-11-13

3
Known defects: who said it, when

From official documentation and published literature, each with the quote and date. We add no assessment of our own.

Coverage

Counts include sub-domains; a dataset can sit in several.

How a dataset is checked

LOINC

  1. 1

    Capture

    • s02loinc.org2026-11-12 · 33 excerpts
    • s03loinc.org2026-11-13 · 7 excerpts
  2. 2

    Quote

    • LOINC is available freely under an open license. It is free for use in both commercial an…Terms of use page
    • The Group 1 Artifacts and Group 1 Documents are Copyright © Regenstrief Institute, Inc. a…Terms of use page
    • The Group 2 Artifacts, Group 2 Documents, Group 3 Artifacts, and Group 3 Documents are Co…Terms of use page
  3. 3

    Apply the rules

    • Rule 1 · counts as a data license: the license field of the dataset's hosting page, or a statement in the docs aimed at the data.

      Terms of use page
  4. 4

    Decide

    RestrictedLink only

    Restricted › link only

    License restricted, judged by public rules and checked by a second-pass model. Link only.

Built for the data supply chain of medical AI

Model labs

Training data that stands up to audit

Provenance is becoming a release gate.

How it works

Research institutions

Less time spent locating data

Biomedical datasets are scattered across dozens of platforms, each describing licenses and sources its own way.

How it works

Pharma

License due diligence for R&D data, done once

AI-driven R&D draws on dozens of databases with different terms: free for academia, licensed for commerce, share-alike, per-institution.

How it works

One configuration line connects an agent

One MCP server and one HTTP API. API keys are self-service in the console. Every answer carries license_status last_verified card_url.

Functions and examplesCreate an API key

{ "mcpServers": { "mdexa": {
    "url": "https://mcp.mdexa.com/mcp" } } }

Example address ahead of launch.

Physician-reviewed data

In preparation · no samples yet
  1. Two physicians, independently

    Each item is answered by two reviewers who do not see each other's work.

  2. Agreement on record

    Agreement is measured and published with the data; disagreements are adjudicated.

  3. Reviewer qualifications

    Licence, specialty and years in practice recorded for every reviewer.

Register a data need

Principles and standards

License rules

  1. Read restrictions broadly, permissions strictly

    Calling a restricted dataset permissive is the costliest mistake.

  2. A repository license is not a data license

    Unless the repository holds only data.

  3. Second-hand evidence never decides alone

    Mirrors and search snippets are listed and labelled.

Public standards

ICD-11
Disease classification reference; disease links switch on after physician review
GB/T 15657-2021
Classification and codes of TCM diseases and patterns
SPDX
License identifiers on every card
schema.org Dataset
Machine-readable metadata on every card page

Data scope

  • No patient records are hosted.
  • Third-party datasets derived from clinical text are flagged separately: 15 today.

Out of scope

  • Judge the efficacy of any medical system
  • Serve files before a human sign-off
  • Build connectors into private enterprise systems

Research

All
  1. Technical reportA Two-Axis Record of License and Provenance for Biomedical Datasets: Schema, Adjudication Rules, and Evidence Requirements

    License and provenance information for biomedical datasets is scattered across license files, hosting-page metadata, documentation, papers and upstream sources, and these sources can disagree.

  2. Technical reportTerms of Use of Public Databases and Terminologies in Drug Discovery: A Structured Survey

    Pharmaceutical and biotechnology users of public drug-discovery databases and coding systems must locate usage terms that are spread over home pages, documentation pages and site-wide policies, and that take very different forms.

  3. Technical reportA Classification of License Identifiers Observed on Biomedical Datasets

    Datasets carry license identifiers of very different kinds: standard open licenses, data-specific agreements, software licenses, model licenses, platform placeholder values, and publisher-specific custom licenses.

Recent changes

  1. Known defects, version notes and access statements from cited sources were added; each shows who said it, when, and what kind of source it…

    54 cards

  2. Re-adjudicated under the current rules; all machine checks still pass.

    53 cards

  3. Added via the automated pipeline and passed all checks.

    12 cards

  4. Practical fields were added by the automated pipeline from stored page snapshots; every statement carries a verbatim quote.

    10 cards

  5. Listed on 2026-11-14. Judged by public rules and checked a second time before listing. The two passes disagreed; the stricter reading is ca…

    10 cards

Next

  1. Roadmap

    Model inference access, model environment containers and a compute environment: planned, not yet available.

  2. Client setup guides

    A copyable setup for each agent client, listed once it passes a real test.

  3. Physician review, first specialty

    Reviewers and specialty to be confirmed.

Plans

Plans

Free

Free

Higher limits, an API key, saved lists and alerts

Sign up
Usage
Query rate120 calls per minute, 5000 per day
API keys1
Call history kept30 days
Saved datasets50
Change alerts10
Tools
Batch list check10

Pro

$120/ year
Founding members $69 / year. First 100 members; the renewal price stays the same.

More headroom for individuals and small teams

ApplyOnline payment is not open yet.
Usage
Query rateUnlimited (fair use)
API keys5
Call history kept180 days
Saved datasetsNo limit
Change alertsNo limit
Tools
Batch list check200

Ultra

$490/ year

The highest limits and the largest batches

Apply
Usage
Query rateUnlimited (fair use)
API keys20
Call history kept365 days
Saved datasetsNo limit
Change alertsNo limit
Tools
Batch list check2,000

No sign-up needed to call the database, up to 20 calls per minute

Full refund within 14 days of activation if you have not used the membership benefits.