Single tool (getCep) with basic but incomplete definition. Tool name starts with verb (good), but description is minimal (21 chars). Input schema present with CEP parameter, but parameter description is weak and contains an example value (59154310) which violates LLM best practice. No output schema documented. No error handling guidance. Tool is marked READ_ONLY, but lacks idempotentHint or readOnlyHint annotations. Spring AI annotation-based approach (@Tool, @ToolParam) is properly used but constraints are missing.

• Expand tool description to 80-150 chars: 'Look up Brazilian address details by postal code (CEP). Returns street, neighborhood, city, state, and additional geographic info. Call this to geocode or validate addresses.'
• Rewrite parameter description without example: 'Brazilian postal code (CEP): 8 digits, with or without hyphen. E.g., format 12345-678 or 12345678.'
• Document CepResponse output schema in the tool or in method-level docs: each field name, type (string), and meaning (e.g., 'logradouro: street name', 'bairro: neighborhood', 'localidade: city name', 'uf: state abbreviation').
• Add error handling: wrap ViaCepClient.getCep() in try-catch and return structured error responses, e.g., 'CEP format invalid' (user-fixable, retryable), 'CEP not found' (retryable), or 'ViaCEP service unavailable' (retryable with backoff).
• Add readOnlyHint=true and idempotentHint=true to @Tool annotation if Spring AI supports it, or document these properties separately for MCP clients.