Single tool 'aivis-speech-synthesis' has a well-structured input schema with 11 parameters, all typed and constrained (numeric ranges, defaults). However, the tool description is in Japanese ('AivisSpeech音声合成のMCPサーバー') and extremely brief (~20 chars in English equivalent). Parameter descriptions are present but minimal. No output schema is documented, the tool returns audioData as ArrayBuffer but this is not formally specified. The tool performs a stateful multi-step operation (audio_query → synthesis) but lacks idempotency hints or confirmation patterns for a destructive audio operation. Error handling exists but lacks recovery guidance for LLMs.

• Expand tool description to 80-150 English characters explaining: 'Synthesize speech from text using AivisSpeech API. Supports speaker selection, voice style, speed, pitch, and intonation control. Returns audio data as WAV format.'
• Document output schema explicitly: { audioData: { type: 'string', description: 'Base64-encoded WAV audio data' }, duration_ms?: number, sample_rate: number }
• Add idempotentHint: false and destructiveHint: true annotations to signal that repeated calls with identical params may produce different audio (non-deterministic synthesis) and have side effects (file I/O, playback).
• Clarify speaker_id and style_id defaults: 'If omitted, uses speaker 0 (default voice). Call get_speakers() first to list available speakers and styles.'
• Translate all error messages and descriptions to English for LLM compatibility. Use structured error responses: { error: 'speaker_not_found', message: 'Speaker ID 999 not found. Available speakers: 0, 1, 2', recovery: 'Call get_speakers() to list valid IDs.' }