Senior Life Care MCP has 22 tools with significant structural gaps. All tools have descriptions (in Korean, describing the service category), and all have input schemas with enum constraints. However, the naming convention is inconsistent (mix of check_, find_, dementia_self_check), parameters lack English descriptions in many cases, and output schemas are not documented. The server appears to be a static data lookup service returning templated responses rather than real integrations. No tools have error handling guidance, recovery hints, or confirmation patterns for sensitive operations. Schemas are present but minimal, no documentation of what each tool actually returns to the agent. This is typical of early-stage informational MCP servers that prioritize breadth over depth.

• Document output schema for every tool. Define what fields each response includes (e.g., facility_id, name, address, phone, availability_status) so LLMs can chain results into downstream calls.
• Translate all parameter descriptions to English or provide English glossaries. LLMs reasoning in English cannot reliably parse Korean descriptions. Use JSON Schema 'description' fields with clear English text.
• Add pagination/limit parameters to find_* tools (find_care_service, find_nursing_home, find_senior_jobs, find_senior_club). Specify 'Returns up to 20 results. Use limit and offset for pagination.' Prevent context overflow.
• Implement error handling and recovery hints. E.g., find_care_service should return 'No services found in region "서울". Try a broader region (e.g., "경기도") or call check_care_eligibility() first.' This guides LLM self-correction.
• Rename dementia_self_check to check_dementia_risk or start_dementia_screening for clarity. Update description to 'Conduct a KDSQ-P (Korean Dementia Screening Questionnaire) assessment. Returns a risk score (0-15) and recommended follow-up (normal, mild suspicion, moderate, severe). Provide 15 yes/no answers as input array.'
• For check_inheritance_info and check_elder_abuse, add required input parameters: region (for jurisdiction-specific guidance), language_preference (optional), or remove them if they are truly no-input reference tools, then explicitly state 'This tool always returns the same content; no input required.'
• Add @readOnlyHint to all tools (all are currently marked READ_ONLY risk, which is correct). Ensure tool annotations are emitted in the MCP response so agents understand they cannot have side effects.

Inferred effective spec: <=2025-11-25.

• Relies on Stateful initialize / Mcp-Session-Id (removed; protocol is stateless) - make each request self-contained