Two tools with comprehensive schemas and detailed parameter descriptions. speak_text has 16 well-typed parameters with constraints (enums, min/max, length limits); play_audio is simpler but properly defined. Descriptions are adequate (60-90 chars) but could be more LLM-optimized. No output schemas documented. Error handling guidance absent. Security considerations present (path sanitization, CLI injection prevention) but not surfaced in descriptions.

• Document output schema for speak_text: return object with {filePath: string, duration?: number, provider: string, model: string}. This enables agents to chain calls and extract audio paths.
• Add error handling guidance to descriptions: 'On API key error, check GEMINI_API_KEY env var. On unsupported voice, call with a different voice from the enum. On disk full, check SPEAK_MCP_OUTPUT_DIR permissions.'
• Expand speak_text description to 150-200 chars: explain it generates speech audio files, when to use Gemini (cloud, no setup) vs Irodori (local, offline), and that it can play audio immediately or save for later.
• Add conditional parameter documentation: 'When provider=irodori, mode, profile, refWav, numSteps, seconds, seed, modelDevice, codecDevice, modelPrecision, codecPrecision apply. When provider=gemini, only voice and stylePrompt apply.'
• Document play_audio return value: {success: boolean, filePath: string, duration?: number, message?: string} so agents know if playback succeeded and can handle failures.