The 'classify' tool is inferred from package.json metadata only. No explicit MCP tool registration, schema, or parameter definitions are visible in the provided source code. The description exists in package.json but is generic marketing copy (97 chars), not an LLM-optimized tool description. No input schema, parameter types, or output schema documentation found. Tool definition cannot be verified from actual code.

• Create explicit MCP tool registration with full JSON Schema input definition. Document all parameters (text, labels, confidence_threshold, etc.) with types and descriptions.
• Rewrite tool description as LLM-optimized prose: 'Classify text into user-defined categories using zero-shot learning. Accepts up to 1,000 lines per request. Returns predicted label and calibrated confidence score (0-1). No API key required.' (target 100-150 chars).
• Document output schema: { label: string, confidence: number (0-1), alternatives?: [{label: string, confidence: number}] }.
• Add parameter descriptions: 'text' (string, required, max 1000 lines), 'labels' (array of strings, required, 2-50 labels), 'confidence_threshold' (number 0-1, optional, default 0.5).
• Add error handling guidance: 'If text exceeds 1,000 lines, split into batches and call classify multiple times. If confidence < threshold, consider requesting user clarification or returning top-2 alternatives.'