MCP Gmail has 11 tools with basic naming (verb_noun pattern) and short descriptions (10-50 chars), but lacks depth. Tool descriptions are minimal, e.g., 'List email categories' and 'List all Gmail labels' do not explain WHEN to use them or what they return. No visible input schemas in the provided source code; tool definitions are exported from individual files but schemas are not shown. Parameters are not documented in the provided excerpts. Output structures are not documented. Error handling and recovery guidance are absent. The server exports tool factories but actual schema definitions, parameter descriptions, and output documentation are not visible in the provided code sample, making it impossible to verify compliance with pattern:tool-description and pattern:constrained-input.

• Expand all tool descriptions to 50-200 characters. Include WHAT the tool does, WHEN to use it (vs similar tools), and WHAT it returns. Example: 'Search for emails using Gmail query syntax (e.g., from:[email protected], has:attachment). Returns up to 20 matching message IDs and subjects. Use message-get to fetch full content.'
• Add explicit parameter descriptions to every tool. For message-search, document 'query' as 'Gmail query syntax string (e.g., from:[email protected], is:unread, has:attachment). See query-syntax prompt for examples.' For message-send, document 'to', 'subject', 'body' with format constraints.
• Document output schemas for all tools. E.g., message-search returns {messages: [{id, subject, from, date}], total_count, next_cursor}. message-get returns {id, subject, from, to, body, labels, date}. This enables LLMs to chain tools correctly.
• Add error handling guidance. For message-send, document: 'Returns error if recipient email is invalid or user lacks send permission. LLM should ask user to verify email address.' For label-delete, document: 'Cannot delete system labels (All Mail, Sent Mail, etc.). Returns list of protected labels if attempted.'
• Implement confirmation pattern for destructive tools. message-move-to-trash and message-send should support a dry_run parameter or return a confirmation step (result: 'input_required') before executing. This prevents accidental data loss.