Open-source MCP server that reads text aloud locally on Windows, macOS, and Linux without API keys or cloud services.
Single tool 'speak_text' has a clear verb-noun name and explicit input schema with proper constraints (minLength=1, maxLength=50000). Description is concise (67 chars) and actionable. Tool annotations present (readOnlyHint=false, destructiveHint=false, idempotentHint=false). However, output schema is undocumented (returns plain string with no structured format), and error handling lacks recovery guidance. Server instructions are strong but tool-level error messages are generic.
Read text aloud locally and return after playback finishes.
Output schema undocumented. Tool returns a plain string ('Completed text-to-speech playback.') with no structured format. LLMs cannot extract metadata (duration, success status, error details) for downstream reasoning.
Error messages lack recovery guidance. ValueError and RuntimeError are caught and re-raised but do not suggest next steps (e.g., 'Queue full, wait 5 seconds and retry' or 'Text exceeds 50k chars, split into smaller chunks').
No pagination or batch support. Single-tool server with no composition patterns. If future versions add multiple tools, lack of batch variants will force sequential calls.
Parameter description 'Do not include secrets or hidden reasoning' is a constraint but lacks format/length guidance in the description itself. Relies on minLength/maxLength in schema, which LLMs may not parse from JSON Schema.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |