This server has a fundamental architectural flaw: it registers a PROMPT, not a TOOL. The @mcp_server.prompt() decorator creates a prompt template, not an executable tool. The categorise_transactions function returns PromptMessage objects, not actionable results. LLMs cannot invoke this as a tool, they can only retrieve the prompt template. The input schema exists (transaction_folder: string) but is never used for tool execution. No actual transaction processing logic is visible. This is a prompt-generation server masquerading as a tool server, making it unsuitable for agent automation.

• Convert @mcp_server.prompt() to @mcp_server.tool() and implement actual transaction processing logic that reads files, categorizes transactions, and returns structured results.
• Define and return a structured output schema: {expenses: [{date, amount, description, category}], income: [{date, amount, description, category}], summary: {total_expenses, total_income}}
• Rewrite description as: 'Categorize expense and income transactions from files in a folder. Accepts folder path, reads transaction files (CSV/JSON), categorizes each into predefined categories, returns structured results with per-transaction success/failure. Requires readable directory with valid transaction files.'
• Add parameter description for transaction_folder: 'Path to folder containing transaction files (CSV or JSON format). Must be readable and contain at least one valid transaction file. Supports relative and absolute paths.'
• Implement error handling: return {success: false, error: 'Directory not found', suggestion: 'Verify path exists and is readable'} for missing folders; {success: false, error: 'No valid files found', suggestion: 'Ensure files are CSV or JSON format'} for empty/invalid files.