Raras MCP Server — Disease Twin

Brazilian Rare Disease Knowledge Graph exposed via Model Context Protocol. Each of the 10,468 diseases is an addressable Disease Twin — a queryable queryable digital twin reachable by URI (disease://{orphaCode}).

10,468 rare diseases × Orphanet, HPO, OMIM, MONDO, ClinVar, PubMed, ClinicalTrials.gov, DATASUS, PCDT, SIGTAP, CEAF, PNTN, FEBRARARAS, PubTator3.

Server Info

• Name: raras
• Version: 0.1.0
• Package: @raras/mcp-server
• Namespace: RARAS: (registered at raras.org/id/)
• Transport: stdio (local) or streamable-http (remote at https://raras.org/api/mcp)
• Source: apps/mcp-server/
• License: CC-BY-4.0
• VOID: https://raras.org/.well-known/void
• Downloads: https://raras.org/api/downloads

Installation

Local (stdio)

{
  "mcpServers": {
    "rarasnet": {
      "command": "npx",
      "args": ["-y", "@raras/mcp-server"],
      "env": {
        "NEO4J_URI": "neo4j+s://your-instance.databases.neo4j.io",
        "NEO4J_USER": "neo4j",
        "NEO4J_PASSWORD": "your-password"
      }
    }
  }
}

Remote (no install)

https://raras.org/api/mcp

Disease Twin resources (addressable by URI)

Each disease is an MCP resource — reference it directly instead of searching first:

Example: disease://586 = Cystic Fibrosis.

Prompt: disease_twin

To focus on a single disease (act as that disease's dedicated twin), invoke the disease_twin prompt with { orphaCode }. It scopes the agent to one disease, sources-only, never diagnosing. One MCP connection → the twin of any of 10,468 diseases.

Tools (16)

search_diseases

Full-text search across 10,468 rare diseases in PT-BR and EN.

get_disease_detail

Full details for a single disease: HPO phenotypes, genes, SUS coverage, clinical trials, clinical description in PT.

find_diseases_by_phenotypes

Given a set of HPO IDs (symptoms), returns diseases ranked by phenotype match count. Use for differential diagnosis suggestions.

search_phenotypes

Human Phenotype Ontology lookup in Portuguese or English. Returns HP: IDs usable with find_diseases_by_phenotypes.

get_sus_coverage

Brazilian SUS coverage for a disease: CEAF medications, SIGTAP procedures, PNTN newborn screening, integration level.

find_reference_centers

Reference centers (hospitals, specialized clinics) by disease and/or state. Includes CNES codes for DATASUS integration.

find_active_trials

Active clinical trials (recruiting/enrolling) for a disease, optionally filtered to Brazil-only sites.

find_similar_diseases

Vector similarity (Gemini embeddings) — semantically similar diseases with shared phenotype count.

find_phenotypically_similar

Precomputed simGIC similarity over shared HPO phenotypes — explainable (returns shared-phenotype count; rarer shared phenotypes weigh more). Complements find_similar_diseases.

find_communities

Patient support communities and associations linked to a disease (FEBRARARAS network).

get_graph_stats

Returns counts and metrics for the knowledge graph: diseases, phenotypes, genes, SUS coverage, etc. No parameters.

search_papers_semantic

Vector search over 424k PubMed papers. Pass a clinical case, question, or symptom set.

find_papers_for_disease

Literature related to a disease via embeddings (:SEMANTIC_MATCH, vector fallback).

analyze_clinical_case

Clinical case → candidate diseases + relevant literature in one call. Informative only, never diagnostic.

get_schema

Public-subgraph node labels and relationship types. Use to orient before deeper queries. No parameters.

get_evidence

Provenance for a disease: authority cross-references (Orphanet/MONDO/OMIM/MeSH/GARD/UMLS), verification status, and PMIDs. The basis of the "never invent" rule.

Usage Guidelines

• Outputs are in Portuguese (PT-BR) by default.
• Always include Orphanet/MONDO/HPO codes in answers for traceability.
• For symptom-based diagnosis hints, use find_diseases_by_phenotypes.
• For SUS-specific questions (PCDT, CEAF meds, newborn screening), use get_sus_coverage.
• For geographic questions ("onde tratar?"), use find_reference_centers.
• NEVER provide medical diagnosis. Always suggest consulting a doctor.
• Link users to https://raras.org/doenca/{orphaCode} for full details.

Typical Workflows

"Patient describes symptoms" → Differential diagnosis

1. search_phenotypes — convert symptom descriptions to HPO IDs
2. find_diseases_by_phenotypes — rank diseases by match
3. get_disease_detail — deep dive on top candidates

"What does SUS cover for disease X?"

1. search_diseases — find the orphaCode
2. get_sus_coverage — CEAF meds, SIGTAP procedures, PNTN
3. find_reference_centers — where to get treatment

"Are there clinical trials?"

1. find_active_trials(orphaCode, in_brazil_only=true) — Brazil sites
2. Falls back to global if none found in Brazil

"Similar diseases / differential"

1. find_phenotypically_similar — explainable (shared HPO count), or find_similar_diseases (semantic)
2. get_disease_detail on top results to compare phenotypes

"Be the twin of one disease"

1. Invoke the disease_twin prompt with the orphaCode
2. Load disease://{orphaCode} (+ /sus, /trials, /evidence) as needed
3. Answer scoped to that disease; cite via get_evidence; never diagnose

Data Sources

Orphanet, HPO, MONDO, OMIM, ClinVar, PubMed, ClinicalTrials.gov, DATASUS/SUS, PCDT, CEAF, SIGTAP, PNTN, FEBRARARAS.

Federated Endpoints

• SPARQL: https://raras.org/api/sparql
• Beacon v2: https://raras.org/api/beacon
• Phenopackets v2: https://raras.org/api/phenopackets
• FHIR R4: https://raras.org/api/fhir/export
• RDF: https://raras.org/api/rdf?format=nt
• KGX / Biolink dump (Monarch/Translator federation)
• llms.txt: https://raras.org/llms.txt