Server has solid naming conventions (verb_noun pattern), comprehensive parameter schemas with types and descriptions, and explicit tool annotations for read/write semantics. However, descriptions are in Russian, limiting LLM comprehension in English-first environments. Output schemas are partially documented (JSON for calendar, text wrapping for email). Error handling via @errors.friendly decorator is present but opaque to callers. Missing: explicit output schema documentation, pagination support for search results, and actionable error messages. Tool composition is clean (8 single-responsibility tools), but email/calendar chains lack intermediate ID documentation.

• Translate all tool descriptions and parameter descriptions to English. Use clear, concise language (50 - 200 chars per description) optimized for LLM comprehension. Example: 'send_email: Sends an email from the user's account. IRREVERSIBLE, requires explicit user confirmation before execution. Parameters: to (recipient emails), subject, body, cc/bcc (optional), attachments (file paths, max 18 MB), reply_to_message_id (optional, to continue a thread).'
• Document output schemas explicitly in each tool's description or as a separate schema field. Example for list_events: 'Returns JSON array of {id, summary, start (ISO 8601), end (ISO 8601), attendees (email list), meet (Google Meet URL)}.'
• Add pagination to search_emails: introduce offset/cursor parameter and return {results: [...], total_count: N, has_more: bool}. Limit default max_results to 10 but allow up to 100.
• Replace @errors.friendly with explicit error classification. Return structured errors: {error: 'invalid_recipient', message: 'Email address invalid: john@. Did you mean [email protected]?', retryable: false}. Guide LLM on next steps.
• Implement a confirm_send_email tool (dry-run variant) that previews recipients, subject, body, and attachments without sending. Agents call this first, then send_email after user approval. Same pattern for create_event.