Claude Can't Watch Videos. This Plugin Fixes That.

What it does

This plugin turns Claude Code into a video viewer by running an MCP server that extracts frames via ffmpeg and audio via your choice of backend—Gemini, OpenAI, or local Whisper. It feeds Claude base64 images and timestamped transcriptions, acting strictly as a perception layer while the model does the actual interpretation. Point it at a local file or YouTube URL, ask a question in plain English, and it adjusts resolution, frame rate, and time window to match.

The interesting bit

Instead of blindly dumping frames, the plugin parses your prompt like a cameraman adjusting lens and shutter speed—high-res screenshots for a single second, or a low-fps skim of an hour-long lecture. For YouTube, it ranks transcript sources—preferring manual subtitles, then auto-captions, then audio transcription—so Claude knows how much to trust the text.

Key highlights

• Plugs into Claude Code via MCP; no separate interface to manage
• Three audio backends: cloud APIs or fully local whisper.cpp for offline use
• YouTube support via yt-dlp with transcript provenance labeled by quality
• Adaptive ffmpeg parameters driven by natural language prompts
• Perception-only architecture: extracts and formats, leaving analysis to Claude

Caveats

• As of v1.0.0, tested only on macOS Apple Silicon with the local Whisper backend
• Requires ffmpeg, Node.js 20+, and optionally yt-dlp installed separately

Verdict

Worth a look if you want to reason about video without leaving Claude Code. Skip it if you need a standalone video analyzer or battle-tested cross-platform stability.