Server has 9 well-named tools with complete input schemas and descriptions. Naming follows verb_noun convention (lookup_cnpj, list_filiais, search_companies). Descriptions are substantive (100-200 chars), explaining what each tool does and key parameters. All parameters have types and descriptions. However, output schemas are not documented, LLMs cannot see what fields to expect from responses. Error handling is minimal; no recovery guidance or actionable error messages. No tool annotations (readOnlyHint, etc.). Parameter constraints are described in text but not formalized as enums or JSON Schema patterns. Missing documentation of pagination behavior and result limits in tool descriptions.

• Document output schemas for all 9 tools. For lookup_cnpj, specify: returns object with razao_social (string), capital_social (number), natureza_juridica (string), simples (boolean), socios (array), estabelecimentos (array with situacao_cadastral, endereco, CNAEs). Use JSON Schema or a structured format in the tool description.
• Add tool annotations to the MCP schema. Mark all tools with readOnlyHint: true since they only query data and never modify state. This signals to agents that these tools are safe to call speculatively.
• Formalize parameter constraints in JSON Schema. Use minLength/maxLength for strings (e.g., query minLength: 3), minimum/maximum for numbers (e.g., per_page maximum: 50), and pattern for CEP (^\d{8}$). Remove constraint descriptions from text and rely on schema.
• Add error recovery hints to tool descriptions. E.g., 'If CNPJ not found, try search_companies() with the company name or partial CNPJ digits.' This guides agents on fallback strategies.
• Document pagination in list_filiais and search_companies descriptions. Specify: 'Returns paginated results. Include page and per_page in request. Response includes total_count and results array. Iterate page until results array is empty or page * per_page >= total_count.'