The server implements a single TTS tool 'speak' with a structured JSON Schema that includes input types and descriptions. However, the tool definition reveals critical gaps: the description is adequate (103 chars, within 10-1024 range), but several parameters lack clear guidance on constraints, expected formats, and interdependencies. The schema is visible and typed, but parameter descriptions are terse and do not explain when to use which provider or voice, nor do they document mutual exclusivity (e.g., provider-specific parameters like 'stability' and 'similarity_boost' that only apply to ElevenLabs). Error handling documentation is absent from the source provided, no recovery guidance for invalid provider, missing voice IDs, or API failures. The tool performs a write operation (speaks/saves audio) but lacks documentation about idempotency, confirmation patterns, or file-save side effects.

• Expand parameter 'voice' description to: 'Voice identifier for the selected provider. For say: use system voice names (e.g., Alex, Victoria). For google: use BCP-47 language code and voice name (e.g., en-US-Neural2-A). For openai: use alloy, echo, fable, onyx, shimmer, or nova. For elevenlabs: use voice ID (e.g., 21m00Tcm4TlvDq8ikWAM).' This matches the constraint and prevents hallucinated values.
• Add 'availability' field to parameter descriptions to clarify dependencies: 'provider=say: rate only. provider=google or openai: model only. provider=elevenlabs: stability and similarity_boost only.'
• Document output schema: 'On success, returns {success: true, message: string, file_path?: string} if saved. If only played: {success: true, message: string}. On error: {success: false, error: string, recovery_hint?: string}'.
• Add recovery guidance in tool description: 'If provider is unavailable, try a different provider (e.g., if google fails, try openai). If voice ID is invalid, call list_voices() to discover valid options. If file save fails, check outputDir permissions.'
• Implement separate 'list_voices' tool accepting provider enum, returning {provider: string, voices: [{id: string, name: string, language: string}]}. This enables LLMs to discover valid voices dynamically instead of guessing.
• Clarify write semantics: 'This tool is idempotent, calling speak() with identical input produces identical output (same file if saved, same audio if played). Safe to retry on transient failures. Not safe to retry if the underlying provider has side effects (e.g., API quota usage).'

Inferred effective spec: <=2025-11-25.

• Relies on Roots (deprecated) - pass files/dirs via tool parameters, resource URIs, or configuration