Loading ToolBench

Reading the current corpus…

video-extract-mcp MCP server · D · ToolBench

video-extract-mcp

github.com/synovial-lionfish82/video-extract-mcp

MCP server that turns any video URL into a transcript and a small set of important, deduplicated keyframes — locally, with no cloud APIs. Supports YouTube, TikTok, Facebook, WeChat Channels, direct MP4/HLS and generic sites.

2026-07-28+STDIO2 tools

53/100

Definition Quality58

Protocol Readiness50

Supportability45

Spec posture

Inferred effective spec 2026-07-28+·current baseline 2026-07-28

Static source inference · medium confidence

Current-spec patterns detected

Summary

Two tools with well-structured schemas and detailed descriptions. resolve_video and analyze_video both have comprehensive parameter documentation and clear intent. However, descriptions lack LLM-optimized brevity (both exceed 200 chars), and output schemas are not explicitly documented in the visible code. Parameter descriptions are strong but could be more concise. No tool annotations (readOnlyHint/destructiveHint) present despite both tools having side effects (WRITE risk).

Tools (2)

analyze_videowrite50/100

Analyzes a video (local file or URL) to extract transcript and keyframes. Performs speech recognition, scene detection, frame selection, and transcript alignment.

resolve_videowrite50/100

Resolves a video URL to extract metadata and optionally download the video file. Supports YouTube, TikTok, Facebook, WeChat Channels, direct MP4/HLS and generic sites via yt-dlp.

Issues (5)

MEDIUMtool-description ↗

Tool descriptions exceed 200 characters and lack LLM-optimized brevity. resolve_video (165 chars) and analyze_video (180 chars) are verbose; optimal range is 50-150 chars for clarity.

resolve_videoanalyze_video

HIGHresponse-shaper ↗

No output schemas documented. Code shows AnalyzeItemResult interface but no explicit return type documentation visible in tool registration. LLMs cannot plan downstream calls without knowing response structure.

resolve_videoanalyze_video

HIGHtool ↗

Missing tool annotations despite destructive operations. Both tools modify state (download files, write transcripts) but lack destructiveHint or idempotentHint annotations required by current MCP spec.

resolve_videoanalyze_video

MEDIUMconstrained-input ↗

Parameter 'frames' enum values ('key', 'even', 'none') lack descriptions explaining what each mode does. LLMs cannot reason about which to select without understanding the difference.

Recommendations

Add explicit output schema documentation to both tools. Document AnalyzeItemResult fields (status, title, duration, frameCount, framePaths, transcript, transcriptPath, manifestPath, warnings) and resolve_video's return structure in tool descriptions or via structured response examples.
Condense tool descriptions to 50-150 characters. E.g., resolve_video: 'Download and extract metadata from video URLs (YouTube, TikTok, Facebook, direct MP4/HLS).' Expand context in parameter descriptions instead.
Add tool annotations: mark both tools with destructiveHint=true (they write files), and analyze_video with idempotentHint=true if re-running with same inputs produces same output.
Document enum values inline. For 'frames' parameter, change description to: 'Frame extraction mode: "key" (scene-detected keyframes), "even" (uniform sampling), "none" (skip frames).'
Add error handling guidance. E.g., resolve_video: 'Returns error if URL is invalid or unsupported. Supports YouTube, TikTok, Facebook, WeChat, MP4, HLS, and generic sites via yt-dlp.' Include recovery hints like 'If URL fails, verify it is publicly accessible.'

Score history

Overall score trend

First recorded score · v2 rubric

53/100

Scored	Grade	Overall	Spec posture	Rubric
2026-09-23	D	53	2026-07-28+	v2