This server has 2 tools with partial schema definitions and moderate description quality. Both tools have input schemas with basic type information, but parameter descriptions are sparse and lack actionability. The schema for search_papers includes a 'default' field in the JSON schema, which is good; extract_info has a minimal schema. Tool descriptions exist but are verbose (containing the Args/Returns breakdown inline rather than as separate response schema documentation). Neither tool description explicitly states whether it modifies state. Missing: output schemas, parameter constraints (e.g., max_results bounds), error handling guidance, and field-level descriptions. The server uses fastmcp (Python), and tool definitions are visible in research_server.py, but without seeing the actual implementation details (error handling, return structures), scores are conservative.

• Add explicit output schema for both tools. For search_papers, document: returns array of objects with fields [paper_id: string, title: string, authors: array, summary: string]. For extract_info, document: returns object {paper_id, title, authors, summary, arxiv_url, ...} or error object {error: string, suggestions: array}.
• Specify parameter bounds: max_results must be 1 - 100 (per arXiv API limits, if applicable). Add a 'topic' constraint: 'Free-form search string, 1 - 500 characters, e.g., "machine learning federated learning"'.
• Rewrite tool descriptions to follow the LLM-optimized pattern: 'Searches arXiv for papers matching a topic and stores results locally. Use this to find research papers before analyzing them. Returns array of paper IDs and metadata. Supports pagination via max_results (default 5).'
• Add error handling guidance to extract_info: 'If paper not found, the tool returns an error. Suggest the user search for papers matching the topic first using search_papers, or clarify the paper_id format (must be valid arXiv ID).'
• Mark search_papers with risk=WRITE in the tool definition comment, and update description: 'WRITE: Modifies local state. Stores paper metadata to disk so subsequent calls to extract_info can find it.'
• Consider splitting extract_info into two tools or enriching search_papers return: (1) search_papers returns detailed metadata directly, or (2) extract_info accepts either paper_id or (topic + index) to avoid external dependency on prior search state.