Claude won't take a video file. But hand it the right frames and the transcript, and it analyzes video better than most "video AI" tools — here's the workflow.
Try to attach an .mp4 to Claude and it's refused — video isn't a supported input. Paste a YouTube link and Claude can at best fetch the page text. So "Claude, analyze this video" fails not because Claude can't reason about video, but because nothing visual ever reaches it.
Claude is, however, excellent at reading images and text. Which means the job is translation: turn the video into the frames that matter plus a transcript, and Claude does the rest.
claude-real-video (crv) is a free, MIT-licensed tool (1.2k+ GitHub stars) built exactly for this. It detects scene changes (instead of blindly sampling every second), deduplicates repeated shots, and transcribes the audio — all on your machine, no upload.
# install (macOS; Windows: winget install Gyan.FFmpeg) brew install ffmpeg pip install "claude-real-video[whisper]" # run on a YouTube / Instagram / TikTok link or a local file crv "https://www.youtube.com/watch?v=..." → crv-out/frames/*.jpg + transcript.txt + MANIFEST.txt
Drag the frames and transcript.txt into a claude.ai conversation and ask:
These are keyframes + transcript extracted from a video. Summarize it, outline its structure, and point out what happens visually that the words don't say.
If you use Claude Code, you can skip the dragging entirely. Install the tool as a skill and Claude watches videos on its own — you just paste a link:
pip install "claude-real-video[whisper]" git clone https://github.com/HUANGCHIHHUNGLeo/claude-real-video.git mkdir -p ~/.claude/skills && cp -r claude-real-video/skills/claude-real-video ~/.claude/skills/
Restart Claude Code, paste a video URL, ask "what's this video about?" — Claude runs the extraction, opens the frames one by one, reads the transcript, and answers. This is the exact flow I use to break down reference videos and study how other creators script theirs.
--grid — tile consecutive keyframes into 3x3 contact sheets, so Claude follows motion as a sequence (and you send 9x fewer images)--why "find the pricing strategy" — written into the manifest so Claude analyzes with that lens instead of a generic summary--whisper-model turbo — sharper transcripts, close to large-v2 accuracy at ~8x the speed--kb ~/notes — save the analysis as a dated note in your own knowledge baseA stack of stills drops the two things that make video video: motion and pacing. Claude can't tell a slow push-in from a handheld chase, or a snappy edit from a lingering one, from disconnected images.
crv Pro ($19 one-time) measures those directly and writes them into the same manifest as plain text: per-shot camera-move classification (static / pan / tilt / zoom / handheld), editing rhythm (shot list, cuts per minute), and a timestamped perception timeline — gestures, expressions, voice pitch shifts, emotion, sound events. Ask Claude why an edit feels fast, and it answers with numbers.
Not as a file — video isn't a supported attachment. Extract keyframes + transcript with crv and Claude analyzes the visuals genuinely well.
No. Extraction and transcription run locally. Only what you paste into Claude afterwards goes to Anthropic.
Yes — same output, any LLM. See the companion guide: how to make ChatGPT watch a video.
The core tool is free and MIT-licensed. The optional Pro add-on for camera/pacing/emotion analysis is $19 one-time.
Free and open source. Works as a Claude Code skill.
Get it on GitHub See crv Pro