MCP server that provides a chatbot interface using IBM Watsonx.ai for medical symptom assessment, with support for greeting resources and symptom assessment prompts.
The server exposes only one tool ('chat') with a minimal input schema and a generic description. The tool lacks structured output documentation, error handling guidance, and comprehensive parameter descriptions. While the code is functional, it does not meet production-grade definition quality standards. The prompt and resource implementations are present but not evaluated here as tools. Parameter documentation is sparse, and output schemas are not formally declared.
MCP tool: generate a chatbot response via Watsonx.ai
Tool 'chat' lacks structured output schema documentation. Returns a free-form string with no declared field structure, making it difficult for LLMs to reason about what data is available or how to extract specific information.
Description for 'chat' is under 100 characters and lacks critical context: when to use this tool vs alternatives, what format the response takes, whether errors are expected, and how to interpret failures. LLM selection suffers from insufficient guidance.
Input parameter 'query' has a minimal description ('User's input message'). No guidance on expected format, length limits, supported languages, or special syntax. LLMs cannot infer constraints.
No error handling or recovery guidance in tool definition or docstring. When inference fails (API timeout, invalid credentials, quota exceeded), the error message is raw: 'Error generating response: {e}'. LLMs cannot distinguish retryable from fatal errors.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 35 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 34 | - | v1 |
No idempotency or side-effect declaration. The tool calls Watsonx.ai which logs activity, consumes quota, and may have rate limits. The description does not clarify whether repeated calls with the same input produce identical output or different responses due to model sampling.
No pagination, result limits, or structured output for multi-turn conversations. If the LLM chains multiple chat calls, there is no mechanism to retrieve conversation history, manage context windows, or limit response length. Single-turn only.