Loading ToolBench
MCP server for zero-shot voice cloning with expressive phoneme synthesis and natural prosody rendering
Critical failures across all dimensions. The single tool 'execute_skill_action' has a generic, action-agnostic name that violates verb_noun convention. Description is 127 chars but lacks WHAT/WHEN/WHY context required for LLM selection. Input schema visible but parameters lack descriptions entirely, 'reference_voice_sample_sec' and 'emotional_tone' have no guidance on valid ranges, enum values, or format. No output schema documented anywhere. The mcp_server.py shows only a stub that prints JSON; no actual tool registration, request handling, or error handling logic is visible. Tool definition appears inferred from client.py rather than explicitly registered in the MCP server. No error handling, no validation, no recovery guidance.
Execute voice synthesis skill action for zero-shot voice cloning with expressive phoneme synthesis
Tool name 'execute_skill_action' is generic and action-agnostic. Does not start with a specific verb (synthesize, clone, generate). LLMs cannot infer intent from the name alone.
Input parameters 'reference_voice_sample_sec' and 'emotional_tone' have NO descriptions. LLMs cannot determine valid ranges, enum values, or format. 'emotional_tone' example shows 'ENGAGING_NARRATOR' but no enum constraint or validation guidance.
No output schema documented. client.py shows a mock response with fields like 'voice_synthesis_id', 'audio_duration_seconds', 'voice_similarity_score_pct', but the MCP server does not declare these in a schema. LLMs cannot plan downstream calls or extract required fields.
mcp_server.py is a stub that only prints a JSON status message. No actual request handling, tool invocation logic, or error handling is implemented. Tool definition appears inferred from client.py rather than explicitly registered.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | F | 31 | 2026-07-28+ | v2 |
No error handling or recovery guidance. If synthesis fails, the LLM receives no actionable error message, no retry classification, and no alternative suggestions.