mcp-getweb demonstrates solid definition quality with complete input schemas across all 8 tools and generally clear descriptions. All tools follow a consistent verb_noun naming pattern (search, fetch) and include structured JSON schemas with proper types and constraints. Tool annotations are present for several tools (readOnlyHint, destructiveHint where applicable). However, output schemas are not documented, the server returns results but does not formally specify the structure of success/error responses, which violates the pattern:response-shaper requirement. Parameter descriptions are present and mostly actionable, but several parameters lack nuance (e.g., 'stream' in felo-search is minimally documented). Error handling is present but generic, errors do not guide recovery with next steps. Some tools (url-metadata) have sparse descriptions ("Extract metadata from a URL (title, description, etc.)") that fall below the 50-character baseline for clarity.

• Add formal output schema documentation for each tool. Define the structure of success responses (e.g., for duckduckgo-search, an array of {title: string, url: string, snippet: string} objects). Use a pattern like 'Returns: array of {"title": string, "url": string, "snippet": string}' in tool descriptions or via separate schema field.
• Enhance error messages with recovery guidance. Instead of 'Search failed: connection timeout', return 'Search failed: connection timeout. Retry with a simpler query (fewer keywords) or try a different search provider (e.g., duckduckgo-search instead of google-search).'
• Expand url-metadata description to clarify its distinction: 'Extract metadata (title, description, author, publish date, image) from a single URL. Use this when you need structured metadata for link previews or citation. For page content, use fetch-url (text extraction) or jina-reader (LLM-friendly markdown).'
• Document felo-search streaming behavior: 'When stream=true, results are returned incrementally as they become available (useful for long-running queries). When stream=false (default), results are batched and returned once. Use streaming for user-facing displays; use batched for batch processing.'
• Add guidance on result set sizes in search tool descriptions. E.g., for google-search: 'Limit to 5-10 results for agent reasoning. Larger result sets dilute signal and increase context overhead. Use pagination if you need to iterate.'
• Validate all parameter constraints in code and return clear error messages. For google-search, if num_results > 10, return 'Invalid num_results: 15. Maximum is 10.' instead of silently capping or returning a generic error.

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry