Server defines 4 tools with schemas and descriptions present. Naming follows verb_noun pattern (ai-text-detection, ai-image-detection, plagiarism-detection, text-compare). Descriptions are adequate (100-200 chars) but lack actionable context on when to use each tool vs. alternatives. Input schemas use Zod with type constraints and descriptions. Critical gap: output schemas are completely undocumented, LLMs cannot predict response structure for downstream chaining. Error handling returns generic text messages without recovery guidance. Tool annotations present (price metadata) but incomplete, missing readOnlyHint/destructiveHint/idempotentHint. Parameters are well-constrained (min/max lengths, URL format validation) but lack enum alternatives where applicable (e.g., language/country codes in plagiarism-detection).

• Document output schemas for all tools. Define response structure (e.g., ai-text-detection returns {score: number, is_ai: boolean, sentences: array}). Include field types and descriptions so LLMs can chain results.
• Replace generic error messages with actionable recovery guidance. E.g., 'Text too short (250 chars). Minimum 300 required. Provide longer text or use ai-image-detection for images.' Include error classification (retryable vs. user-fixable).
• Add readOnlyHint: true to all tool annotations. This signals to agents that these tools are safe to call without side effects and can be retried freely.
• Expand text-compare description to 100+ chars. Explain when to use it (e.g., 'Compare two texts to measure similarity. Use after plagiarism-detection to verify specific passages.'). Include similarity score range and interpretation.
• Convert language and country parameters to enums. Define valid ISO 639-1 language codes (en, fr, es, de, it, pt, nl, tl, pl, id) and ISO 3166-1 country codes. Prevents LLM hallucination of invalid values.