Server has 7 tools with basic schemas and descriptions, but significant gaps limit production readiness. All tools have descriptions (10-200 chars, mostly adequate), and input schemas are present with type definitions. However, output schemas are entirely undocumented, LLMs cannot predict response structure for planning. Parameter descriptions lack constraint details (ranges, enums, formats). Error handling is minimal: tools catch exceptions but return raw error strings without recovery guidance. No tool annotations (readOnlyHint, idempotentHint). Naming is verb-forward (search_, get_, find_) but some tools combine multiple concerns (research_topic does search+citations+bibtex+translate in one call, violating single-responsibility). No pagination support despite tools returning lists. Missing validation of inputs (e.g., source enum not enforced, count bounds not checked).

• Document output schemas for all tools. Define what fields each tool returns (e.g., search_papers returns array of {title, authors, date, pdf_url, id}). Use JSON Schema or explicit type descriptions so LLMs can plan downstream calls.
• Add enum constraints to parameters. Change 'source' param to enum: ['arxiv', 'semantic', 'both']. Add min/max to 'count' (e.g., 1-50). Enforce in code with early validation and clear error messages.
• Split research_topic into separate tools: search_research_papers, get_paper_citations, generate_bibtex, translate_text. Each tool does one thing; agent composes them. Enables reuse and clearer error handling.
• Enhance error messages with recovery guidance. Instead of 'Could not fetch paper: <exception>', return 'Paper not found. Try search_papers() to find the paper ID first, or verify the DOI format (e.g., 10.1038/nature12373).'
• Add pagination to list-returning tools. Add 'limit' (default 5, max 50) and 'offset' (default 0) parameters. Return 'total_count' and 'has_more' in response. Cap results at 20-50 items per call to avoid context bloat.