Model Context Protocol (MCP) server exposing FHIR repository search, GraphRAG knowledge graph queries, and medical data visualization tools for AI-powered medical chat applications.
Server has 11 tools with complete input schemas and descriptions, but quality is inconsistent. All tools are READ_ONLY with no state-modifying operations, which simplifies error handling but limits utility. Descriptions are present but generic (avg ~120 chars, below the 194-char baseline). Parameter descriptions lack specificity about formats, ranges, and constraints. No output schemas documented. Tools follow verb_noun naming but lack composition guidance, many visualization tools (plot_*) return Plotly-compatible data without specifying the exact structure. No error handling patterns visible. STDIO transport caps protocol readiness at 50.
Retrieve full details of a FHIR DocumentReference resource. Decodes hex-encoded clinical notes and returns complete document with metadata and full text content.
Traverse knowledge graph relationships from seed entities. Performs multi-hop graph traversal to find related entities and discovers semantic connections between medical concepts.
Get knowledge graph statistics including entity counts by type, relationship counts, and high-confidence entities. Useful for understanding the scope and quality of the knowledge graph.
Combined FHIR + GraphRAG search using Reciprocal Rank Fusion. Searches both FHIR document text and knowledge graph entities, then fuses results for optimal ranking. Best for complex medical queries.
Generate interactive pie/bar chart data showing distribution of entity types (SYMPTOM, CONDITION, MEDICATION, etc.). Returns Plotly-compatible chart data.
Output schemas not documented. Tools return 'Plotly-compatible chart data' and 'network graph data' without specifying field names, types, or structure. LLMs cannot plan downstream processing or extract specific fields.
Parameter descriptions lack format/constraint details. 'max_depth' accepts '1-3 hops' but no validation guidance. 'chart_type' accepts 'pie' or 'bar' but not declared as enum. 'granularity' accepts 'day', 'week', 'month' but not constrained.
No error handling patterns documented. Tools do not specify what happens on invalid input (e.g., nonexistent FHIR ID, invalid entity text), whether errors are retryable, or what recovery steps the LLM should take.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | 2026-07-28+ | v2 |
Generate network graph data showing entity relationships for visualization. Returns nodes and edges in format compatible with network visualization libraries.
Generate timeline visualization showing patient encounters over time. Returns Plotly-compatible timeline chart data with dates and event counts.
Generate bar chart showing most frequent symptoms in the knowledge graph. Returns top N symptoms with their occurrence counts.
Search FHIR DocumentReference resources by text content. Decodes hex-encoded clinical notes and performs full-text search. Returns matching documents with clinical note previews.
Search medical knowledge graph for entities matching query terms. Returns entities with types (SYMPTOM, CONDITION, MEDICATION, etc.) and confidence scores, plus documents containing those entities.
Visualize GraphRAG search results showing the query, entities, relationships, and source documents. Returns structured data for network visualization.
Tool composition unclear. Multiple search tools (search_fhir_documents, search_knowledge_graph, hybrid_search) exist without guidance on when to use each. No documentation of which tool output fields chain into downstream tool parameters.
Visualization tools lack pagination/limits. 'plot_entity_network' accepts 'max_nodes' (default 50) but no guidance on what happens if the knowledge graph has millions of entities. No total count or next_cursor returned.