An MCP server for fetching and transcribing YouTube videos
Three tools with basic verb-noun naming (get_, extract_, search_) but significant quality gaps. All tools have descriptions (10-60 chars, below the 194-char baseline), and input schemas are present with type definitions. However, descriptions are too brief to guide LLM selection effectively. No output schemas documented. No error handling guidance. No tool annotations (readOnlyHint). Parameters lack detailed constraints and format specifications. The server conflates two distinct concerns (YouTube API transcript retrieval vs. manual audio extraction) in separate tools without clear composition guidance.
Extract and transcribe audio from a YouTube video when no transcript is available
Retrieve the transcript from a YouTube video
Search for YouTube videos and return the top results
Descriptions are too brief (25-60 chars vs. 194-char baseline). 'Retrieve the transcript from a YouTube video' lacks WHEN to use it, prerequisites, or what it returns. LLMs cannot distinguish get_transcript from extract_transcript without richer context.
No output schemas documented. Callers cannot know what fields to expect (e.g., does get_transcript return transcript_text, transcript_language, source, or all three?). This forces LLMs to guess and breaks tool chaining.
Parameter 'language' lacks enum constraint. Description says '(en or vi)' but no formal enum. LLMs may pass 'english', 'English', 'ENG', or other variants, causing silent failures.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 51 | <=2025-11-25 | v2 |
No error handling guidance. Flask code returns 404 or 500 with generic messages ('Failed to get transcript'). LLMs receive no recovery hints (e.g., 'Try search_youtube_video first' or 'Video may be private or deleted').
No tool annotations. All three tools are read-only (no state modification), but readOnlyHint is absent. This prevents agents from reasoning about safety and retry logic.