A Node.js MCP server for video search, analysis, and transcription that integrates with the VidNavigator Developer API
VidNavigator presents 6 tools with complete input schemas and descriptions for all parameters. Tool names follow verb_noun conventions (search_videos, analyze_video, get_video_transcript, answer_followup_question, get_usage, transcribe_video). However, descriptions are generic (average 145 chars, below the 194 baseline) and lack LLM optimization. Descriptions state WHAT the tool does but rarely explain WHEN to use it or provide dependency hints. Parameters have types and descriptions, meeting minimum standards, but lack actionable constraint guidance. Output schemas are entirely undocumented, LLMs cannot predict what fields to extract or chain to downstream tools. Error handling is present (ErrorCode, McpError imported) but no recovery guidance or error categorization visible in tool definitions. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite all tools being READ_ONLY. Security: API_KEY is injected via environment variable (correct pattern), but no audit logging visible. Composition: each tool is single-purpose, which is correct. No batch variants or multi-step composition patterns. The naming 'answer_followup_question' is less clear than 'ask_video_question', it implies context from a prior interaction that may not be obvious to LLMs cold-starting.
Analyze a video's content and generate summary, extract key information
Answer a follow-up question about a previously analyzed video
Get current usage statistics including API calls, storage usage, and subscription limits
Get the transcript for a specific video without AI analysis
Search for videos using advanced AI analysis with prompt engineering, content analysis, and relevance ranking
Transcribe non-YouTube videos (Instagram, TikTok, Facebook, etc.)
Output schemas entirely undocumented. LLMs cannot infer response structure, cannot extract chaining IDs (e.g., if search_videos returns video_id, can it be passed to analyze_video?), cannot plan downstream calls.
Descriptions are generic and lack LLM-optimization. Average 68 chars (vs 194 baseline). Do not explain WHEN to use each tool or distinguish overlapping ones (e.g., analyze_video vs get_video_transcript, transcribe_video vs get_video_transcript). Descriptions rarely provide dependency hints or recovery guidance.
No tool annotations despite clear read-only semantics. All 6 tools should carry readOnlyHint=true per current MCP spec. This guides LLMs and clients on retry safety and concurrency.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 54 | 2026-07-28+ | v2 |
| 2026-03-09 | D | 55 | - | v1 |
Parameter constraints lack actionable guidance. 'language' in transcribe_video says 'Optional ISO2 language code (en, fr, es, etc.)', includes examples that LLMs may reuse literally. Should use enum for supported languages or regex pattern (^[a-z]{2}$) with explicit list in description.
Tool composition ambiguity. 'answer_followup_question' implies prior context (previous analyze_video call) not captured in the schema. LLMs may not understand the dependency. Description should explicitly state 'Requires: video must have been previously analyzed via analyze_video().'
No error handling guidance in tool definitions. Code imports McpError and ErrorCode but tool descriptions do not indicate what errors are retryable, user-fixable, or fatal. For example, 'invalid video_url', is this user error (fix the URL) or fatal?
Pagination and result limits undocumented. search_videos accepts max_results (1-10, default 5) but description says 'note: API uses default limit', unclear what the server's actual limit is. Does search_videos support cursor-based pagination for >10 results? Response should document total_results and next_cursor if applicable.