Paper Search MCP has 11 well-named tools with consistent verb_noun patterns (search_*, get_*). All tools have descriptions (avg ~80 chars, within baseline 34-392). Input schemas are present with Zod validation and type constraints (min/max bounds on numeric params). However, output schemas are NOT documented, responses are JSON.stringify'd without explicit field documentation, forcing LLMs to infer structure. Error handling wraps all tools but returns generic JSON error objects without recovery guidance. Parameter descriptions are adequate but lack dependency hints (e.g., 'year' format varies by source). No tool annotations (readOnlyHint, idempotentHint) despite all being read-only operations.

• Document output schemas for all tools. Specify the structure of paper objects returned (e.g., {title: string, authors: string[], year: number, doi?: string, url?: string, abstract?: string}). This enables LLMs to extract and chain data correctly.
• Enhance error handling with actionable recovery guidance. Instead of '{error: "Unknown error"}', return '{error: "API timeout", retryable: true, suggestion: "Try again in 5 seconds or use an alternative source like Semantic Scholar"}'. Categorize errors as retryable, user-fixable, or fatal.
• Add tool annotations. Set readOnlyHint: true on all 11 tools to signal they are safe, non-destructive operations. This helps agents reason about side effects and retry strategies.
• Standardize year parameter format across all tools. Document as 'ISO year range (e.g., 2024, 2020-2024, 2020-, -2024)' or use a single canonical format. Add regex pattern validation in Zod schemas.
• Add pagination support with offset/limit or cursor-based parameters. Document max_results as a per-page limit and provide next_cursor in responses so agents can fetch additional results without re-querying.