Async FastMCP video transcription service using yt-dlp with local GPU (Parakeet + pyannote 4.0) and AssemblyAI fallback
Server has 9 tools with basic naming and descriptions, but significant gaps in schema documentation and parameter detail. Tool names follow verb_noun convention (transcribe, search, list_transcripts, get_metadata, get_comments). Descriptions are present but brief (10-60 chars), below the 50-200 char LLM-optimized baseline. Input schemas visible for all tools with type declarations, but parameter descriptions are minimal or missing context. No output schemas documented. Error handling and recovery guidance absent. Tool composition is reasonable (single responsibility), but parameter constraints (enums, ranges, defaults) are underdeclared. Security considerations (API keys, rate limits) not evident in visible code.
Get comments for a video.
Get full metadata for a video.
Get the status of a transcription job.
List available transcripts.
Read a transcript by video ID. Returns full speaker-diarized markdown.
Search transcripts by content.
Queue a video for transcription.
Descriptions are too brief (10-60 chars vs. 50-200 char baseline). LLMs cannot determine when to select tools or understand prerequisites. E.g., 'Queue a video for transcription' lacks context on async behavior, expected latency, or failure modes.
No output schemas documented. LLMs cannot plan downstream tool calls or extract required fields (e.g., job_id from transcribe, video_id from search). Responses are opaque.
Parameter descriptions missing or minimal. 'limit' appears in 4 tools with identical description 'Maximum number of results (default: X)' but no range constraints (e.g., 1-100). 'sort' in get_comments lacks enum declaration.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2026-07-28+ | v2 |
Queue all videos in a playlist for transcription.
Search YouTube for videos.
No error handling or recovery guidance visible. Tools lack documentation on failure modes (e.g., invalid URL, transcription timeout, rate limits). LLMs cannot self-correct or retry intelligently.
Tool composition: 'transcribe' and 'transcribe_playlist' are separate, but no guidance on when to use each. 'search' (transcripts) vs 'yt_search' (YouTube) are confusingly similar names. LLMs may conflate them.