Single tool 'scrape' has a clear action-verb name and explicit input schema with proper types and descriptions. However, the output schema is not documented, the tool returns raw JSON from the UseScraper API without specifying what fields to expect. Error handling returns HTTP error details but lacks actionable recovery guidance. Missing tool annotations (readOnlyHint/destructiveHint/idempotentHint) despite being a read-only operation. Parameter descriptions are adequate but could be more prescriptive about constraints.

• Document the output schema for the 'scrape' tool. Example: 'Returns an object with fields: title (string), content (string), metadata (object with url, extracted_date, format).' Make this precise to what UseScraper actually returns.
• Add readOnlyHint annotation to the tool definition in ListToolsRequest: {..., readOnlyHint: true}. This marks the tool as safe for automatic execution without confirmation.
• Enhance error handling to return categorized, actionable messages. For 'Invalid URL' → 'Invalid URL format. Ensure it starts with http:// or https://'. For API auth failures → 'API key invalid or expired. Check USESCRAPER_API_KEY environment variable.'
• Specify the structure of 'extract_object' parameter with an example and valid keys. If it maps to UseScraper's extraction API, document what properties are accepted (e.g. {"type": "object", "properties": {"css_selectors": {...}}}).
• Add description of size limits: 'Pages larger than 1MB are truncated to the first 50KB of text content. For large pages, use format=text to reduce size.' This prevents LLM confusion when responses seem incomplete.
• Consider adding 'timeout_seconds' parameter (optional, default 30) to let agents control how long to wait for slow pages, improving resilience to hangs.