MongoDB MCP server has 12 tools with basic schemas and descriptions, but significant gaps prevent higher scores. All tools have descriptions (10-80 chars, mostly adequate), and input schemas are present with type definitions. However, output schemas are completely undocumented, LLMs cannot predict what fields to expect from responses, breaking tool chaining. Parameter descriptions are minimal (e.g., 'Database name' lacks format/constraint details). No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels. Error handling is present but generic, no recovery guidance. Naming is verb-first and clear (list_, create_, drop_, etc.), which is good. The server implements security features (permissions, audit logging, rate limiting) but these are not exposed in tool definitions, leaving LLMs unaware of constraints.

• Document output schemas for all 12 tools. Example: list_databases should return {databases: [{name: string, sizeOnDisk: number}], totalSize: number}. This enables LLMs to chain tools and extract correct fields.
• Add tool annotations to all tools. Use readOnlyHint: true for list_*, *_stats, list_indexes. Use destructiveHint: true for drop_* tools. Use idempotentHint: true for create_* and rename_* if they are idempotent.
• Enhance parameter descriptions with constraints. E.g., 'Database name (1-64 chars, alphanumeric + underscore, no spaces)' instead of just 'Database name'. Add format/pattern constraints to schema.
• Add error recovery guidance. When a database is not found, return: 'Database "mydb" not found. Available databases: [list]. Call list_databases() to discover valid names.'
• Implement dry-run mode for destructive operations. Add optional 'dryRun: boolean' parameter to drop_database and drop_collection. Return impact summary (collections affected, size to be deleted) without executing.