The arXiv MCP server defines 3 search/retrieval tools with complete input schemas and reasonable descriptions. All tools follow verb_noun naming (search, search_advanced, get_paper). Parameter schemas are well-typed with descriptions. However, output schemas are not documented in the source code, the server returns Pydantic model instances (SearchResult, Paper) but the actual field structures are not visible in the provided code snippet. Descriptions are adequate (70-150 chars) but lack explicit guidance on when to use each tool vs. alternatives. The search_advanced tool requires 'at least one search field' validation (seen in code), but no error recovery hints are provided. Tool annotations are present (readOnlyHint, openWorldHint), which is good for protocol compliance. No destructive operations, so confirmation patterns not needed.

• Document the complete output schema for SearchResult and Paper. Include example JSON or a reference to the Pydantic model definition. Show LLMs exactly what fields to expect: id_arxiv, title, abstract, authors (array of strings), categories, url_abstract, url_pdf, date_published, etc.
• Add 'At least one of title, abstract, author, category, or id_arxiv must be provided' to the search_advanced description, or restructure the parameters to make at least one required in the schema definition.
• Convert sort_by to an explicit enum in the input schema: {"enum": ["relevance", "date_desc", "date_asc"], "default": "relevance"} instead of a plain string.
• Create a companion discovery tool: list_arxiv_categories() → returns all valid categories (e.g., {categories: ['cs.AI', 'cs.CL', 'cs.LG', 'stat.ML', ...]}). Call this first to guide category filtering.
• Wrap error responses with actionable recovery hints. Instead of raising ValueError('arXiv returned HTTP 503...'), catch it and return a structured error response that tells the LLM: 'Temporary service error. Retry in 60 seconds, or try a simpler query.'
• Clarify get_paper's id_or_url parameter: add constraint description like 'Pass either an arXiv ID (format: YYMM.NNNNN, e.g., 2301.00001) or a full URL (https://arxiv.org/abs/...). The tool will extract and use the ID.'
• Add pagination guidance to search and search_advanced descriptions: 'Results are paginated. Use page and page_size to iterate. The response includes total_results to guide agent loops.' This helps LLMs understand when to fetch next pages.