ResearchPaperMCP has 5 tools with basic descriptions and partial schemas. All tools have descriptions (10-200 chars), but parameter descriptions are sparse or missing. Input schemas are visible for all tools with type declarations, but lack depth in constraints and validation guidance. No output schemas documented. Error handling is minimal, most failures return raw exception strings rather than actionable recovery guidance. Tool naming follows verb_noun convention (search_web, fetch_web_content, academic_research, get_paper_id, find_related_papers), which is good. However, parameter descriptions are often single-phrase and lack format/constraint details. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) present. Composition is reasonable, tools chain logically (get_paper_id → find_related_papers), but output schemas are not documented, forcing LLMs to infer structure.

• Document output schemas for all tools. E.g., search_web returns [{title: string, link: string, snippet: string}]; academic_research returns [{paperId: string, title: string, authors: [{name: string}], year: int, citationCount: int, abstract: string, ...}]. Include these in tool descriptions or as separate schema documentation.
• Replace raw exception strings with actionable error messages. E.g., 'Error accessing page: Connection timeout' → 'Failed to fetch URL (timeout after 15s). Verify the URL is accessible and try again.' Include retry guidance for transient errors.
• Add min/max bounds to numeric parameters. E.g., 'limit: integer, minimum 1, maximum 20 (default 5)' for academic_research. Add format guidance to string params: 'query: string, 1-500 characters, e.g., "attention mechanism transformer"'.
• Add tool annotations to schema. Mark all tools with readOnlyHint: true since they perform no mutations. This signals to LLMs that these tools are safe to call speculatively.
• Implement pagination for search_web and academic_research. Add offset/page and total_count to responses. Current implementation caps results at 20 but does not expose pagination metadata, limiting agent ability to fetch more results if needed.