MCP Video Analyzer

What guimatheus92/mcp-video-analyzer MCP server does

The guimatheus92/mcp-video-analyzer MCP server turns video sources into structured information that an AI agent can inspect. Inputs can include supported platform URLs, direct video links such as .mp4 or .webm, Loom links, and local video files. Results can combine speech transcripts, extracted frames, OCR text, a time-aligned timeline, and video metadata.

The full analysis operation can also include comments, chapters, and an AI summary when those details are available from the source. For narrower tasks, the server exposes separate operations for transcripts, metadata, frames, individual timestamps, and short frame bursts. This lets an agent request only the data needed for a particular question.

How it works

The guimatheus92/mcp-video-analyzer MCP server exposes eight MCP tools:

• analyze_video returns a complete analysis.
• analyze_videos processes multiple sources and produces one structured result per source, with resumable batch processing.
• get_transcript uses native captions when available and can fall back to Whisper.
• get_metadata retrieves metadata, comments, and chapters without downloading the video.
• get_frames extracts key frames using scene changes or dense one-frame-per-second sampling.
• analyze_moment examines a selected time range using burst frames, transcript content, and OCR.
• get_frame_at returns one frame at a requested timestamp.
• get_frame_burst samples several frames from a narrow interval for motion or animation inspection.

Frame extraction can use yt-dlp and the bundled ffmpeg-static package. Chrome or Chromium provides an optional fallback. Static videos without scene changes can use uniform temporal sampling so frame and OCR results are still available.

Setup and configuration

Run the server with Node.js 22.12 or newer. The package starts as an MCP stdio server with npx mcp-video-analyzer@latest; the README also provides an executable npx -y mcp-video-analyzer@latest form for command-line use. MCP clients can register the package as a stdio process.

The guimatheus92/mcp-video-analyzer MCP server does not require credentials. Install yt-dlp separately for YouTube, Vimeo, TikTok, Instagram, X, Twitch, Dailymotion, and Facebook URLs. It is optional for direct URLs and local files, and can improve Loom downloads. Without yt-dlp or Chrome, direct files and local files can still use ffmpeg-based frame extraction, while platform URLs may produce an install warning.

The bundled one-shot CLI accepts an input source and can write frame images to a selected output directory. MCP_CACHE_DIR can relocate the persistent cache, and MCP_FRAME_MAX_WIDTH can set a default frame width limit. CLI options also control detail level, frame count, OCR languages, Whisper settings, output fields, refresh behavior, and output location.

Tools and capabilities

Choose get_metadata when downloading frames is unnecessary, get_transcript for speech-only workflows, or get_frame_at and get_frame_burst for targeted visual checks. Use analyze_moment to investigate a known interval, and analyze_videos for multiple sources. The complete tool combines transcript, visual, OCR, timeline, and metadata results in one response.

Limitations and notes

Platform coverage depends on yt-dlp and the source being accessible. Transcription may rely on native captions or the Whisper fallback, and comments, chapters, and AI summaries are available only when supported by the source. Partial processing problems are reported as warnings; hard failures are distinct from those partial results. The server performs local processing and stores extracted frame files in a cache unless an output location is specified.