Single tool 'generate_sfx' has a complete input schema with proper JSON Schema structure and all parameters typed. Description is adequate (145 chars, within 10-1024 range). However, output schema is undocumented, the tool returns a TextContent with a file path, but the LLM has no formal specification of what fields to expect. Parameter descriptions are present but lack actionable constraints (e.g., prompt_influence range 0.0-1.0 is mentioned but not enforced; no guidance on valid duration ranges). Error handling exists but is generic (McpError with code/message) and does not guide recovery. No tool annotations (readOnlyHint, destructiveHint, idempotentHint). The tool modifies state (writes files) but the description does not explicitly warn of side effects.

• Document the output schema explicitly. Define the response as {type: 'object', properties: {file_path: {type: 'string', description: 'Absolute path to generated audio file'}, duration_seconds: {type: 'number'}, ...}}. This lets LLMs know what fields are available.
• Add min/max constraints to numeric parameters. E.g., 'duration_seconds (1-300 seconds, default 5)' and 'prompt_influence (0.0 to 1.0, default 0.75)'. Include these in both the description and JSON Schema minValue/maxValue.
• Revise tool description to declare state modification: 'Generates a sound effect and writes it to disk. This is an irreversible operation, the audio file is created and persisted. Returns the absolute file path.'
• Enhance error messages with recovery guidance. E.g., if ELEVENLABS_API_KEY is missing, return 'API key not configured. Set the ELEVENLABS_API_KEY environment variable and restart the server.' If duration is out of range, return 'Invalid duration: 500 seconds exceeds maximum of 300. Please use a value between 1 and 300.'
• Add tool annotations: set destructiveHint=true (writes files), idempotentHint=false (versioning means repeated calls produce different files). This helps agents understand retry semantics.