Claude Real Video

What claude-real-video does

The claude-real-video MCP server project prepares video material for inspection by Claude, ChatGPT, Gemini, or another model. It does not perform the final LLM analysis itself. Instead, it creates a folder containing selected frames, transcript text, and a manifest that explains the extracted material.

Inputs can be local video files or URLs supported through yt-dlp, including links that require an authorised Netscape cookie file. The tool can also save a full copy of the soundtrack for models that support audio input. A --why prompt records the intended analysis focus in the manifest, while --kb writes the result as a dated Markdown note in a chosen notes directory.

How it works

Video processing is performed locally. URLs are fetched with yt-dlp; local files are copied for processing. ffmpeg then makes one chronological pass that selects scene changes and maintains a frame-density floor. A configurable pixel-difference comparison removes near-duplicates, using a sliding window so that a previously seen shot is not repeatedly emitted after a cutaway.

For text, the tool first checks for an adjacent .srt or .vtt file or an embedded subtitle track. If no subtitles are available, it can use Whisper to transcribe the audio, provided the optional transcription dependencies are installed. Videos without audio are skipped cleanly. The generated MANIFEST.txt brings the frames and text together for downstream model use.

Setup and configuration

Install the core package with pip install claude-real-video. Add the whisper extra when audio transcription is needed. Python 3.10 or newer is required, and ffmpeg plus ffprobe must be installed separately and available on PATH. The project supports macOS, Windows, and Linux.

Run crv with a URL or filename. python -m claude_real_video is an alternative entry point. Important options include --out for the output directory, --scene for scene sensitivity, --fps-floor for minimum sampling density, --max-frames for the frame limit, --lang for Whisper language selection, and --dedup-threshold and --dedup-window for duplicate removal. --no-transcribe, --keep-audio, --report, --why, --kb, and --cookies enable the corresponding optional behaviors.

The claude-real-video MCP server can also be installed as a Claude Code skill by copying the included skill directory into ~/.claude/skills/.

Tools and capabilities

• Extracts keyframes at scene changes instead of relying only on fixed intervals.
• Applies pixel-based deduplication and supports a hard frame cap.
• Produces chronological frames, transcript text, and a manifest.
• Uses existing subtitles before falling back to Whisper transcription.
• Creates contact sheets with ordered frames for model review.
• Saves the original soundtrack with --keep-audio.
• Generates an HTML keep/drop report with --report.
• Exposes a Python process function for programmatic use.

Limitations and notes

The project does not itself send video, frames, or transcripts to an LLM; users choose what to provide to another model. Whisper transcription requires the optional dependency and still relies on ffmpeg. URL access depends on yt-dlp support and, for login-gated sources, a user-supplied authorised cookie file. Re-running the command overwrites the output directory. The included instructions also warn users to download only content they are permitted to access.

The README mentions a separate crv Pro offering for additional video-analysis reports, but the documented core workflow covers local extraction, transcription, manifests, and optional audio.

Getting started with this claude-real-video MCP server

Always refer to the official documentation for the most accurate and up-to-date information.