Raindrop MCP has well-structured tool definitions with consistent naming (verb_noun pattern), clear descriptions (avg ~100 chars), and complete input schemas with types and descriptions. All 17 tools follow the verb_noun convention (list_, get_, create_, update_, delete_, parse_, check_). Schemas are properly typed with enums for constrained inputs (view, sort). However, output schemas are not documented, LLMs cannot predict response structure for chaining. Error handling is minimal; no recovery guidance or actionable error messages. No tool annotations (readOnlyHint, destructiveHint) despite clear risk classifications. Parameter descriptions are adequate but lack format constraints (e.g., 'URL to parse' should specify valid schemes). Missing pagination guidance for list tools despite accepting page/perpage params.

• Document output schemas for all tools. Example: 'Returns Collection object with fields: _id (number), title (string), count (number), public (boolean), view (enum: list|simple|grid|masonry), cover (array of URLs), parent (nested collection ref).'
• Add tool annotations to tool definitions. Mark delete_* tools with destructiveHint=true, read-only tools with readOnlyHint=true, and idempotent tools with idempotentHint=true.
• Implement error handling with recovery guidance. Example: 'Collection not found (404). Verify collection ID with list_collections() or check if it was deleted.'
• Add format constraints to parameter descriptions. Example: 'URL to parse (must start with http:// or https://)', 'Highlight color (hex format: #RRGGBB or named color: red, blue, green).'
• Clarify pagination behavior in list_* tool descriptions. Example: 'Returns paginated results (0-indexed pages, max 50 per page). Response includes total count and items array. Use page parameter to iterate.'