Server has 8 tools with complete input schemas and descriptions. Naming follows verb_noun pattern (search_*, get_*, api_status). However, descriptions embed example values (e.g., 'CGRP migraine treatment', 'NCT06123456') which LLMs tend to reuse literally. Output schemas are not documented, responses are free-text markdown strings rather than structured objects. No error handling guidance (e.g., 'If search fails, try...'), no pagination support for list results, and no tool annotations (readOnlyHint, etc.). Parameter descriptions are adequate but lack format constraints (e.g., 'pmid must be numeric'). Overall: functional but below production baseline.

• Replace free-text markdown responses with structured JSON objects. E.g., search_pubmed should return {articles: [{pmid, title, authors, journal, year, doi, abstract_snippet, url}], total_count, next_cursor} instead of a markdown string.
• Remove example values from descriptions. Replace 'Search keywords (e.g. 'CGRP migraine treatment')' with 'Search keywords (e.g., disease name, drug name, or condition). Supports boolean operators (AND, OR, NOT).'
• Add error handling guidance to tool descriptions. E.g., 'If no results found, try broader keywords or check spelling. If API timeout occurs, retry after 5 seconds.'
• Add pagination to search tools. Include total_count and next_cursor in responses. Accept optional offset/limit parameters (default limit=5, max=50).
• Annotate tools with readOnlyHint=true (all 8 tools are read-only). This signals to agents that retries are safe and no side effects occur.