11 tools with complete schemas and descriptions. Naming is clear and verb-based (list_, search_, read_, mark_, flag_, move_, delete_). Descriptions are adequate (100-200 chars typical) but lack LLM-optimized guidance on WHEN to use each tool vs. alternatives. Parameters are typed and described. Error handling returns structured error dicts with error_type classification, which is good. However, output schemas are not documented, callers must infer structure from code. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels. Batch operations (mark_unread, flag_messages) are present, reducing composition friction. Missing: pagination guidance, per-item success/failure for batch ops, and recovery hints in error messages.

• Document output schemas for all tools. E.g., list_messages returns [{uid: int, subject: str, sender: str, date: ISO8601, unread: bool, flagged: bool}, ...]. Include this in tool descriptions or a separate schema field.
• Add tool annotations: mark read_message, list_messages, search_messages, get_attachment with readOnlyHint=true; mark delete_messages with destructiveHint=true; mark move_messages with reversibleHint=true. Agents will then know which operations are safe to retry.
• Enhance error messages with recovery hints. E.g., FolderNotFoundError should return 'Folder "INBOX" not found. Available folders: [list]. Try list_folders() to discover valid folder names.'
• For batch operations, return per-item results: {successful: [uid1, uid2], failed: [{uid: uid3, error: 'Message not found'}]}. This lets agents retry only failed items without re-processing successes.
• Add pagination guidance to list_messages and search_messages descriptions: 'Returns up to limit messages (default 20). Use limit parameter to fetch more; no cursor support, fetch all and filter client-side if needed.'