What is claude-real-video (crv)?
claude-real-video, shortened to crv, exists because Claude cannot take a video file. It downloads a URL with yt-dlp or reads a local file, extracts a frame at every scene change with ffmpeg (plus at least one per second), drops near-duplicates against a sliding window of recent frames, and transcribes the audio with Whisper unless the video already has subtitles. The output folder holds the frames, a frames.json map of source timestamps, the transcript and a MANIFEST.txt telling the model how to read it.
It ships as a Python package (pip install "claude-real-video[whisper]"), an agent skill installed with npx skills add, a Claude Code plugin, a local web page (crv-web) and an MCP server (crv-mcp) with five tools. The skill and plugin are the main route, so it is reviewed here rather than under MCP servers. Built by Leo Huang, it is MIT-licensed with about 2,280 GitHub stars; version 0.10.7 reached PyPI on 26 September 2026. A paid crv Pro adds camera-motion, emotion and sound-event analysis for a one-time $29.
Is claude-real-video (crv) worth it?
crv solves a real gap with an unglamorous pipeline. Fixed one-frame-per-second sampling floods the context on a static screencast and misses fast cuts; scene-change detection plus deduplication sends each distinct shot once. The README's own example is a 58-second clip where 1 fps gives 58 frames and crv keeps 26, packed into 3 contact sheets with --grid. Every kept frame keeps its source timestamp through deduplication and renaming, so the model can cite 'frame_012 @ 00:03:41' instead of guessing.
The author's measurements give a sense of cost. On a 3-minute 640x360 video on a Mac mini M4, the default mode kept 170 of 180 frames in 23.5 seconds, an estimated 52,000 image tokens at Anthropic's published rate; --max-frames 80 halves that to about 25,000. These are self-reported figures on the author's benchmark file, and they show that a busy video is not cheap: a talking head collapses well, a fast-cut reel does not. get_transcript in the MCP server returns words only, for long talks where frames add little.
Processing is local, which matters for meeting recordings and client footage. Only the frames and text you then give a cloud model leave the machine. A local SQLite index at ~/.crv/memory.db lets crv-ask search across everything already watched; CRV_NO_MEMORY=1 turns it off. The --cookies and --cookies-from-browser options read your browser's login cookies to fetch gated videos, which is convenient and a credential the tool can then use, so they belong on your own accounts only.
The limit is conceptual. The model sees stills and a transcript, not motion or tone; crv Pro's camera, pacing and emotion analysis is the paid answer, and on-screen text recognition is in the $9.90 Codex build but not the free edition. Gemini reads video natively and needs no pipeline, at the cost of uploading the file to Google and sampling at a fixed rate. For Claude users, crv is the most-starred open tool built for this job, with videoagent the alternative when a full agent framework is wanted rather than a preprocessing step.
Where it earns its place
A product manager who missed a 90-minute customer call recording
Find what the customer said about pricing and what was on screen at the time
Run crv on the recording with --from and --to around the screen share and --why 'pricing objections', then ask Claude. Free and local; only the selected frames and transcript go to the model.
A developer learning from a 40-minute conference talk
Get the code shown on the slides, not just the speaker's words
Raise --frame-width to 1600 so slide text stays legible and add --text-anchors if subtitles exist. Expect a larger token bill per frame; get_transcript alone is cheaper when slides are not needed.
A social-media analyst at a small agency reviewing competitor reels
Break down the hooks and cuts in 20 short videos a week
Batch crv over the links with --grid so each reel becomes a few contact sheets, and query across them later with crv-ask. Pacing and shot-type analysis need crv Pro at $29 one-time.
Before you commit
- Setup
- Install ffmpeg, then pip install "claude-real-video[whisper]" (or [fast] for faster-whisper, [mlx] on Apple Silicon) and npx skills add HUANGCHIHHUNGLeo/claude-real-video, or install the Claude Code plugin. For MCP, pip install 'claude-real-video[mcp]' and claude mcp add crv -- crv-mcp. Python 3.10+.
- Integrations
- Skill for Claude Code, Cursor, Codex, Copilot and Gemini CLI; MCP tools watch_video, get_frames, get_transcript, search_memory and list_watched; a Python API (process()); yt-dlp for YouTube, Instagram, TikTok and other sites.
- Data and privacy
- Download, frame extraction and transcription run locally; nothing is sent until you give frames or text to a model. A local SQLite memory index at ~/.crv/memory.db stores transcripts and timestamps (CRV_NO_MEMORY=1 disables it), and MCP analyses are cached in ~/.cache/crv-mcp. Cookie options read your browser login cookies.
- Cost at scale
- Free locally; CPU or Apple GPU time for Whisper. The model bill is image tokens: about 52,000 for 170 frames of a 3-minute video at 640x360 on the author's benchmark, so cap frames on long or busy videos.
- Limits
- Burned-in captions are not detected as subtitles; on-screen text recognition is not in the free edition; motion, tone and emotion require crv Pro. Single maintainer.
What’s good
- Scene-change frames with source timestamps preserved
- Runs locally; video never uploaded
- Skill, plugin, CLI and five-tool MCP server
- Clip windows with --from and --to for long recordings
What’s not
- Fast-cut videos still cost tens of thousands of image tokens
- Stills and words only; motion and tone are in paid Pro
- On-screen text recognition not in the free edition
- Needs ffmpeg installed separately
Best for
Claude Code and other agent users who need a model to answer questions about videos without uploading them.
claude-real-video (crv) alternatives
claude-real-video (crv): common questions
- Is claude-real-video (crv) worth it?
- The sensible way to let Claude reason about a video today: local, cheap on tokens and honest about timestamps, though it shows the model stills and words rather than motion. We rate it 4.0 out of 5.
- How much does claude-real-video (crv) cost?
- Free and open source (MIT). Optional crv Pro is a one-time $29; a separate Codex build is a one-time $9.90.
- What are the downsides of claude-real-video (crv)?
- Fast-cut videos still cost tens of thousands of image tokens; stills and words only; motion and tone are in paid Pro; on-screen text recognition not in the free edition; needs ffmpeg installed separately.
- Who is claude-real-video (crv) best for?
- Claude Code and other agent users who need a model to answer questions about videos without uploading them.
- How much does claude-real-video (crv) cost for a team or in production?
- Free locally; CPU or Apple GPU time for Whisper. The model bill is image tokens: about 52,000 for 170 frames of a 3-minute video at 640x360 on the author's benchmark, so cap frames on long or busy videos.
- Is claude-real-video (crv) safe for company data?
- Download, frame extraction and transcription run locally; nothing is sent until you give frames or text to a model. A local SQLite memory index at ~/.crv/memory.db stores transcripts and timestamps (CRV_NO_MEMORY=1 disables it), and MCP analyses are cached in ~/.cache/crv-mcp. Cookie options read your browser login cookies.
- What are the best claude-real-video (crv) alternatives?
- The closest alternatives we have reviewed are VideoAgent and MarkItDown.
Answers checked Oct 31, 2026.
Make claude-real-video (crv)? Show this rating on your site
