fx-mcp has 10 well-named tools with clear descriptions (avg 85 chars, all >20 chars). All tools use Zod schemas with typed parameters and descriptions. However, output schemas are not formally documented, responses are text-only (no structured JSON fields returned). Error handling is minimal (basic try-catch with error messages). No parameter constraints (enums, ranges) despite many accepting free-form strings (currency codes, token symbols). Missing idempotency hints, permission gates, and audit logging. Tools are read-only and well-composed, but lack production-grade robustness.

• Return structured JSON objects instead of text. E.g. get_fx_rate should return {ecb_rate: number, ecb_date: string, wise_rate: number, wise_fee: number, competitors: [{name, rate, fee, receive}]}. This enables LLMs to extract and chain data.
• Add enum constraints for currency codes and token symbols. Use Zod .enum() or .refine() to validate against ISO 4217 and known token lists. Include validation error messages: 'Invalid currency: got "USDA", did you mean USD?'
• Enhance error handling with recovery guidance. Catch API failures and return: {error: string, retryable: boolean, suggestion: string}. E.g. 'ECB rate unavailable. Try Wise rate or retry in 30s.'
• Add pagination to list_supported_pairs and get_all_fx_rates. Accept limit (default 20, max 100) and offset/cursor parameters. Return {items: [...], total: number, next_cursor: string}.
• Declare tool annotations in server.tool() calls. Add readOnlyHint: true to all tools. This signals to clients that tools are safe to cache and retry without side effects.