AI-driven hospital information system with MCP client for medical data and AI-assisted diagnostics
Scoring was not performed
Input/output schemas not visible in source code for 19 of 22 tools. Schema definitions either missing or inferred from route handlers. No explicit JSON Schema registration visible.
Descriptions universally terse and uninformative (15-25 chars). Do not explain WHEN to use tool, WHAT it returns, or WHAT parameters mean. Examples: 'AI assistant tool for medical diagnosis' (48 chars), 'Get list of available AI providers' (35 chars). No recovery guidance.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 22 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 20 | - | v1 |
Medical domain tools lack constraint documentation. Tools like check-drug-interaction, check-contraindication, check-dosage, and review-prescription have no parameter validation, output format specs, or error handling guidance. Critical for safety-sensitive medical use.
No error classification or recovery guidance in any tool description. Medical AI errors (e.g., dosage validation failure) must be actionable. Current state: no indication of retryable vs fatal errors, no recovery steps.
Tool names lack verb clarity. 'query' (generic, non-descriptive), 'tools' (ambiguous, does this list tools, call tools, or describe tools?), 'tool/call' (unclear if this executes or inspects). Compare against baseline: 90% of A+ tools start with action verb like get_, create_, search_.
Parameter descriptions missing for tools with visible schemas. 'chat' tool has parameters (messages, message, context, maxTokens, temperature, stream) but descriptions are minimal or generic.
No documentation of output schemas. Agent cannot predict what fields to expect from tools. This forces trial-and-error integration and risks context loss on unexpected response structure.