Server has 22 tools with consistent naming (verb_noun pattern) and reasonable descriptions (avg ~80 chars). All tools have input schemas with types and descriptions. However, output schemas are undocumented (responses are text-formatted, not structured JSON), error handling is minimal (generic try-catch with errorResult), and parameter descriptions lack actionable constraints (e.g., date formats, enum values not always explained). No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels. Missing pagination guidance in list tool descriptions. Composition is sound, tools are single-purpose and chainable via IDs.

• Convert all textResult() responses to structured JSON objects with typed fields. E.g., finmap_invoices_list should return {invoices: [{id, number, date, amount, currency, status, counterparty}], total, hasMore} instead of markdown text.
• Add tool annotations to server.tool() calls: readOnlyHint for list/get tools, destructiveHint for delete tools, idempotentHint for create/update tools. This exposes risk levels to LLMs via MCP protocol.
• Enhance error messages with actionable recovery guidance. E.g., catch 404 errors and return 'Invoice not found. Try finmap_invoices_list() to search by date range or status.' Categorize errors as retryable vs user-fixable.
• Add descriptions to reference/lookup tools explaining their purpose: 'Call this first to discover available accounts/categories/suppliers before creating operations. Returns a list of {id, name, ...} for use in other tools.'
• Implement a dry-run parameter for destructive tools (finmap_invoice_delete, finmap_operation_delete) or add a separate confirm_delete tool that requires explicit confirmation before executing the delete.