Strong tool naming (all verb-prefixed, clear intent). Descriptions are present and actionable (avg 150 chars, within 10-1024 baseline). Input schemas fully defined with Zod, including enums, constraints, and descriptions. However, output schemas are not documented, responses are JSON-stringified but structure is opaque to LLMs. Error handling is minimal (generic try-catch). No tool annotations (readOnlyHint, destructiveHint). Missing pagination guidance for search_social_videos despite potentially large result sets.

• Document output schemas for all tools. Add TypeScript interfaces or JSON Schema definitions in comments/docs showing expected response structure (e.g., VideoMetadata { title: string, author: string, stats: { views: number, likes: number }, ... }).
• Add tool annotations via MCP SDK: mark all tools with readOnlyHint=true since they only read public data. This signals to agents that these tools are safe to call without confirmation.
• Enhance error handling: catch specific error types (network, not-found, rate-limit, validation) and return structured errors with actionable guidance. E.g., 'Video not found at URL. Verify the URL is public and the platform is supported (YouTube, TikTok, Instagram, X, Facebook).'
• Add pagination guidance to search_social_videos description: 'Returns up to 100 results. For large result sets, use limit and sort parameters to narrow results. No cursor-based pagination currently supported.'
• Validate and document the 'schema' parameter in extract_from_video. Provide an example JSON Schema in the description or accept a schema_type enum (e.g., 'product_review', 'sentiment_analysis') instead of free-form objects.