MCP server for SaluteSpeech — speech recognition and synthesis (Russia)
Server has 5 well-named tools with clear descriptions and visible Zod schemas. All tools follow verb_noun naming (recognize_*, synthesize_*, list_*, get_*). Descriptions are concise (50 - 150 chars) and explain WHAT each tool does. Schemas are properly typed with enums and constraints. However, parameter descriptions lack actionable detail (e.g., 'language' param doesn't explain when to use each language code; 'voice' param lists speaker names but no guidance on selection). Output schemas are not documented, LLMs cannot predict response structure. Error handling wraps exceptions into isError responses but provides no recovery guidance. No tool annotations (readOnlyHint, idempotentHint). Missing dependency hints and chaining IDs in responses.
Check status of an async SaluteSpeech recognition task by ID.
List available SaluteSpeech models and voices for recognition and synthesis.
Recognize speech from a local audio file. Auto-detects format from extension.
Speech recognition via SaluteSpeech. Accepts Base64 audio, returns text transcription.
Text-to-speech via SaluteSpeech. Accepts text, returns Base64-encoded audio.
Output schemas not documented. LLMs cannot predict response structure (e.g., what fields does list_models return? Is it an array or object? What are the field names?). Forces agents to guess and parse unstructured text.
Parameter descriptions lack actionable constraints. 'language' param lists valid codes (ru-RU, en-US, kk-KZ) but doesn't explain when to use each. 'voice' param lists speaker names but no guidance on voice selection or audio quality differences. 'sample_rate' in recognize_file says 'required for headerless files' but doesn't explain what happens if omitted for WAV files.
No tool annotations. All tools are read-only (no state modification) but lack readOnlyHint annotation. Agents cannot determine which tools are safe to retry or call speculatively without explicit hints.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 69 | 2026-07-28+ | v2 |
Error messages provide no recovery guidance. safeTool() wrapper catches errors and returns isError=true with raw message, but does not suggest next steps. E.g., 'Audio is 2.5 MB, which exceeds the 2 MB limit' (good) but no hint to use async flow. 'Task not found' gives no guidance on valid task IDs or how to list tasks.
recognize_file accepts file_path parameter without validation or sanitization. No path traversal checks visible. If an agent passes '../../../etc/passwd', the tool may attempt to read it. Requires input validation and sandboxing.