MCP server providing FHIR operations and tools for healthcare AI agents with comprehensive patient data access, vital signs analysis, and PDF report generation
This MCP server has severe definition quality issues across nearly all dimensions. While tool names follow verb-first patterns (get_, search_, generate_), the definitions lack proper schema documentation, parameter descriptions, and output structures. Only 6 tools are defined, all in a shared Python module (shared/fhir_tools.py), but the actual schema registration is not visible in the provided source code. The server implements HTTP transport with proper Dockerization, but tool definitions are minimal and would not meet production standards for LLM agent interaction. Critical failures: (1) No visible input schemas in the source code for any tool, only a terse parameter summary is provided; (2) Parameter descriptions are minimal or generic (e.g., 'Patient data including demographics and clinical information' for a complex object); (3) No documented output schemas; (4) No error handling or recovery guidance; (5) Tool compositions suggest multi-step operations bundled into single tools (e.g., generate_comprehensive_assessment takes raw patient_data and assessment_data objects, forcing the agent to assemble these elsewhere).
Generate comprehensive AI assessment report in PDF format
Get detailed encounter information
Get comprehensive patient data for AI assessment
Get FHIR tool configuration
Get vital signs analysis
Search FHIR resources
No visible input schemas in source code. Tools are listed with only parameter type hints (e.g., 'patient_id: string') but no formal JSON Schema definitions. Per hard scoring rule: if a tool has NO input schema at all, its schema score MUST be 0.
Parameter descriptions are missing or trivial. 'Get FHIR tool configuration' (23 chars) does not describe the purpose or when to use get_tool_config. 'FHIR resource type to search' lacks context on valid values and format.
Complex parameters (patient_data, assessment_data, searchParam) lack any schema definition. 'Patient data including demographics and clinical information' does not specify required fields, types, or structure. LLM cannot validate input or construct objects correctly.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 39 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 25 | - | v1 |
No documented output schemas. Tools like get_patient_comprehensive_data return unstructured comprehensive patient data, but the response structure is not documented. LLMs cannot plan downstream calls or extract fields reliably.
No error handling or recovery guidance documented. No mention of what happens if patient_id is invalid, encounter not found, or vital signs analysis fails. No categorization of errors as retryable, user-fixable, or fatal.
generate_comprehensive_assessment bundles multiple concerns: accepts pre-assembled patient_data and assessment_data objects, then generates PDF. This forces the agent to construct complex nested objects before invocation. Should be split: (1) tool to fetch/format patient data, (2) tool to compute assessment, (3) tool to render PDF.
Generic search tool accepts free-form 'searchParam' object. No enum or constraint on valid search parameters per FHIR resource type. No documentation of which resource types are supported or which search parameters are valid for each.
get_vital_signs_analysis has 'days' parameter with default 30, but no documented range or constraints. Can agent pass days=99999? days=-1? No validation rules in description.
Tool definitions are inferred from a brief summary list; actual MCP tool registration code is not visible in the provided source. Per hard scoring rule: if you cannot see explicit tool registration, cap that tool's overall at 50. All tools should be capped at 50, but scoring conservatively at 32-42 based on visible description quality.