Server has 5 tools with complete input schemas and descriptions. Naming follows verb_noun pattern (translate_text, detect_language, etc.). Descriptions are adequate (100-200 chars) but lack actionable context for LLM selection. Parameters are well-typed with enums for language_format. Output schemas are not documented, LLMs cannot predict response structure. Error handling converts gRPC errors to human-readable text but lacks recovery guidance. Missing tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles.

• Document output schemas for all tools. Example: 'Returns {translated_text: string, model_used: string, detected_language?: string}'. LLMs need explicit field names to chain calls.
• Enhance error messages with recovery guidance. Instead of '❌ gRPC error UNAVAILABLE: connection refused', return '❌ Translation service unavailable. Ensure TRANSLATE_GRPC_URL (localhost:5100) is reachable. Retry in 30s or check service logs.'
• Add tool annotations via inputSchema metadata. Mark translate_text/detect_language with readOnlyHint: true; mark synthesize_speech with destructiveHint: true (writes files).
• Expand tool descriptions with selection context. E.g., 'Use translate_text for single strings; use translate_localization for JSON files with multiple keys. Supports models: m2m100_418M, nllb_200_distilled_600M.'
• Return synthesize_speech output file path in response. Include 'file_path' field so agent can immediately reference the generated WAV without directory guessing.