Local MCP + REST server wrapping Whisper large-v3 on GPU: transcribe local media or YouTube URLs.
Strong foundation with 7 well-named tools using clear verb_noun patterns (transcribe_*, analyze_*, extract_frames, estimate). All tools have descriptions (avg 120 chars, within 10-1024 range). Input schemas are complete with proper JSON Schema types and parameter descriptions. However, output schemas are not documented in the source code, LLMs cannot see what fields to expect from responses. Error handling is minimal; no recovery guidance or actionable error messages visible. Tool descriptions lack dependency hints and prerequisites. Risk annotations (READ_ONLY, WRITE) are present but not formalized as toolAnnotations in the MCP protocol.
Analyze a local media file and extract frames at scene changes or uniform intervals.
Analyze a YouTube URL and extract frames at scene changes or uniform intervals.
Estimate token count and cost for transcription or frame analysis.
Extract frames from a media file at scene changes or uniform intervals, with optional time window.
Transcribe a local media file. Accepts container paths (/data/foo.mp4), Windows host paths (D:\whisper-data\foo.mp4), or filenames under DATA_DIR.
Transcribe media uploaded as raw bytes. Local files are streamed to the server as a body upload, so the file does not need to live under DATA_DIR.
Output schemas not documented. LLMs cannot see what fields transcribe_file, analyze_file, or estimate return, forcing them to guess downstream field names and breaking tool chaining.
No error handling guidance. Errors return raw HTTP status codes or stack traces with no recovery hints. E.g., if transcribe_file fails due to unsupported format, the LLM has no guidance on what to try next.
Tool descriptions lack dependency hints and prerequisites. E.g., transcribe_upload does not explain that the file must be streamed as raw bytes in the request body, not as a path parameter.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 68 | <=2025-11-25 | v2 |
Transcribe a YouTube URL (or any yt-dlp-supported URL).
Risk annotations (READ_ONLY, WRITE) are present in metadata but not formalized as MCP toolAnnotations (readOnlyHint, destructiveHint, idempotentHint). LLMs cannot parse these hints from comments.
estimate tool description is vague ('Estimate token count and cost'). Does not explain what 'token count' means (input tokens? output tokens? total?), or clarify that cost is optional and depends on TOKEN_COST_USD_PER_MTOK env var.