videoscrub has 4 well-defined tools with explicit schemas, descriptions, and tool annotations. All tools are READ_ONLY and properly registered via server.registerTool(). Naming is clear (video_info, transcript, frames, motion) and action-oriented. Descriptions are present and substantive (avg ~150 chars). However, several parameters lack descriptions (fps, width in frames tool; start_s/end_s in motion lack detail about their interaction), and output schemas are not formally documented, responses are text or image blobs without structured field definitions. Error handling is minimal; no recovery guidance or actionable error messages visible. Parameter constraints exist (e.g., crop validation, width bounds) but are not fully described in parameter annotations.

• Document output schemas formally. For video_info, return {title: string, duration_s: number, width: number, height: number, fps: number, transcript: {status: string}, chapters: array}. For transcript, return {text: string, lines: [{time_s: number, text: string}], nextOffset?: number}. For frames, return {header: string, sheets: [image_data]}. For motion, return {text: string, buckets: [{time_s: number, luma_change: number}]}.
• Expand parameter descriptions. Add to 'fps': 'Omit to auto-fit up to 1 fps within the range and width. If specified, must be positive.' Add to 'start_s'/'end_s': 'Range in seconds; clamped to [0, duration]. For motion, range must not exceed 600 seconds.'
• Add error recovery hints. When throwing 'empty range' error, suggest 'Ensure start_s < end_s and both are within [0, duration_s].' When frame count exceeds limit, suggest 'Reduce fps, narrow the range, or increase width to fit more frames per sheet.'
• Implement explicit pagination for transcript. Return {text: string, nextOffset?: number, totalLines?: number, pageSize: number} so LLMs know when results are truncated and can fetch more.
• Document crop constraints in parameter description: 'Region in 0 - 1 coordinates from top-left. Must satisfy x + width ≤ 1 and y + height ≤ 1.'