Model Context Protocol server for YouTube API - search videos, manage playlists
The server provides 5 tools with complete input schemas and basic descriptions, but falls short of production-grade quality. Tool naming follows verb_noun convention well (search_videos, create_playlist, etc.), but descriptions lack LLM-optimization depth and omit critical context like prerequisites, side effects, and when to use each tool vs. similar ones. All parameters have type definitions and most have descriptions, but descriptions are minimal (10-30 chars), providing insufficient context for LLM decision-making. No output schemas are documented, only the input schemas are specified. Error handling is basic (generic client-not-initialized check) with no recovery guidance. The server lacks any parameter validation, constraint documentation, or examples of expected behavior. Parameter descriptions do not clarify ranges (e.g., maxResults accepts 1-50 but this is stated in description only, not as schema minItems/maxItems). No security patterns are visible (API keys are injected via environment but no scope declarations or permission gates). The tools are well-composed (one responsibility each) and follow a logical chain (search → create → add), but lack idempotency guarantees and confirmation flows for destructive operations.
Add a video to a playlist
Create a new YouTube playlist
Delete a YouTube playlist
List user playlists
Search for YouTube videos
Tool descriptions are too short and lack context. 'Search for YouTube videos' (30 chars) and 'Create a new YouTube playlist' (30 chars) do not explain when to call each tool, what they return, or prerequisites. LLMs cannot distinguish between similar tools or plan multi-step workflows without fuller descriptions.
No output schema documented. Tools return JSON responses (e.g., search_videos returns videoId, title, description, channel, publishedAt, thumbnail), but this structure is not declared in the tool definition. LLMs cannot plan downstream tool calls or extract fields without documented return types.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 48 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 0 | - | v1 |
Parameter descriptions lack constraint details. 'Max results (1-50)' is stated in text, but not enforced as JSON Schema minProperties/maxProperties or validated in code. No minLength/maxLength, pattern, or enum definitions beyond what's stated in descriptions. LLMs may pass invalid values (e.g., maxResults: 100 or 0).
No error recovery guidance. Error responses are generic ('Error: YouTube client not initialized') with no actionable next steps. LLMs cannot self-correct or decide whether to retry, ask the user, or abort.
Destructive operation (delete_playlist) lacks confirmation or dry-run support. Agents make mistakes, invoking delete_playlist with an incorrect playlistId could permanently delete data without warning or undo path.
No idempotency guarantees documented. If add_to_playlist is called twice with the same (playlistId, videoId), does it fail, silently succeed, or add the video twice? Agents retry on errors, undefined idempotency behavior can cause duplicate entries.
No scope or permission declarations. Tools do not state what OAuth scopes they require (e.g., youtube.readonly vs youtube.manage). This makes it impossible to configure least-privilege agent access or audit permissions.
Parameter 'privacy' in create_playlist has an enum (private, public, unlisted) with a default (private), but no description of what each value means or when to use it. An LLM cannot reason about whether public playlists are safe for sensitive content.
Natural identifiers not supported. All tools accept only system IDs (playlistId, videoId). LLMs cannot search for playlists by human-readable name or filter by channel, every interaction requires opaque IDs, forcing extra lookup calls.