Zendesk MCP has well-structured tool definitions with complete JSON schemas and clear descriptions. All 6 tools have proper naming (verb_noun pattern), documented parameters with types and constraints, and descriptions 50-200 chars. However, output schemas are not documented, LLMs cannot predict response structure. Error handling lacks recovery guidance (no 'try X instead' messages). Parameter descriptions are functional but minimal; some lack context on when to use optional fields. No batch operations despite agents likely calling create/update in loops.

• Document output schemas for all tools. Example: 'Returns {ticket_id: number, subject: string, status: string, created_at: ISO8601, comments: [{id, body, author, created_at}], requester: {id, name, email}, assignee: {id, name, email}}'. This lets LLMs plan downstream calls.
• Add error recovery guidance to descriptions. E.g., zendesk_get_ticket: 'If ticket not found, try zendesk_search_tickets() with the ticket ID or subject.' zendesk_update_ticket: 'Invalid status, must be one of: new, open, pending, hold, solved, closed.'
• Clarify pagination in zendesk_search_tickets: 'Returns up to perPage results. Check response.total_count to determine if more pages exist. Use page parameter to fetch subsequent pages.'
• Document requesterEmail/requesterName behavior in zendesk_create_ticket: 'If requesterEmail matches an existing user, that user is assigned as requester. If not, a new user is created. If omitted, ticket is created without a requester.'
• Consider adding batch variants: zendesk_create_tickets (array of ticket objects) and zendesk_add_ticket_comments (ticketId + array of comments) to reduce sequential calls when agents process multiple items.