OpenAI-compatible TTS MCP server with voice discovery and public audio URLs.
Server defines 2 tools with explicit schemas and descriptions. list_voices has a clear, actionable description (Chinese, but translates to voice discovery with fallback behavior). text_to_speech has comprehensive parameter documentation with type constraints (enums for response_format, numeric bounds for speed). However, descriptions are in Chinese, limiting LLM comprehension in English-first contexts. Output schemas are not formally documented in the code, return types are inferred from TypeScript types (SpeechUrlResult, VoiceListResult) but not exposed in tool registration. Error handling is present (fallback voices, timeout handling) but recovery guidance is minimal. No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles.
获取当前 OAI_TTS_MODEL 的可用音色。优先调用 GET /v1/tts/voices?model=...;上游失败、超时或响应无效时自动返回内置 OpenAI 音色,并在结构化结果中标记 source=fallback。
调用 OpenAI 兼容 POST /v1/audio/speech 生成语音,并上传到 Urusai 文件托管服务。传入 instructions 可指定语气、情绪、节奏或风格;未指定 voice 时使用 OAI_TTS_VOICE;未指定 speed 时默认 1.25。成功后仅返回公开音频 URL。
Tool descriptions are in Chinese; LLMs trained primarily on English will struggle to understand when/why to invoke these tools. Descriptions must be in English for cross-language agent compatibility.
Output schemas (SpeechUrlResult, VoiceListResult) are defined in TypeScript but not exposed in tool registration. LLMs cannot see the structure of returned data and must infer fields from context, risking downstream tool call failures.
text_to_speech is a destructive/write operation (uploads audio, modifies external state) but lacks tool annotations (destructiveHint=true). Agents cannot distinguish safe read-only tools from risky write tools.
Error messages are returned as plain strings in fallback scenarios (e.g., 'provider returned HTTP 404'). No structured error classification (retryable vs. user-fixable vs. fatal) to guide agent recovery logic.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 56 | 2025-06-18+ | v2 |
text_to_speech accepts 'instructions' parameter (up to 4096 chars) for tone/emotion/style but no guidance on which TTS models support this feature. Documentation mentions 'gpt-4o-mini-tts supports best' but this is buried in description, not a formal constraint.