Did our sandbox establish a working video workflow?
Our sandbox passed installation and build commands, but the test command exited 1. It did not verify an interactive Claude Code session or video-answer accuracy.
Our run on September 29, 2026, at commit ac47d1f passed installation and build commands but failed its test command. The build output identified version 1.3.2. The harness recorded 16 seconds for installation, 12 seconds for the build and 14 seconds for the test step.
Our harness stores command success flags, exit codes, elapsed wall-clock times, timeout status and output tails. It stores parsed test summaries separately. The attached stored parser summaries report 138 passed, 10 failed, 148 total and zero skipped tests. These are not verified exhaustive counts of unique tests; no new execution or complete-log reconciliation was performed.
The raw footer names four failing test files. Audio, frame and segment extraction failures include spawn ffmpeg ENOENT; a metadata failure includes spawn ffprobe ENOENT. Those errors show that our sandbox could not find the executables. Separately, the silence-analysis integration test produced no intervals where its assertion expected at least one. The supplied tail does not establish that assertion failure's cause.
The framework footer reports a duration of 1.69 seconds, while the harness records 14 seconds for the command step. These measure different scopes. The stored dependency record lists 224 packages occupying 199 MB; the installation tail separately says it added 207 packages and audited 208. Neither tail figure replaces the stored dependency count or establishes audit coverage.
The MrKeyoor testing methodology explains command recording and sandbox limitations. It describes the publisher's current procedure, not proof that this older run used today's methodology revision.
What does Claude Video Vision give Claude?
Claude Video Vision extracts video frames through ffmpeg and processes audio through Gemini API, local Whisper, or OpenAI API.
The project README describes a perception layer: Claude receives frame images and timestamped transcription, then interprets them. Its six tools cover extraction, structural analysis, closer inspection, metadata, configuration and dependency setup. The documented workflow adjusts frame rate, time range and resolution to the question.
That makes targeted recording review a useful workflow to evaluate. Our run did not measure whether sampling captures a brief UI defect, whether transcripts are accurate, or whether Claude answers correctly. Start with a short recording whose important moment you already know and compare the returned evidence with its timeline.
Which setup requirements apply?
The tested commit requires Node.js 20+; the supplied current README requires Node.js 24+. Our version 1.3.2 result does not validate current version 1.4.0.
Use the README pinned to our tested commit when interpreting the older run. Use the requirements matching your chosen installation when setting up a new evaluation. Both supplied README states require ffmpeg, with yt-dlp needed for YouTube URLs. The current package manifest identifies the newer version and Node requirement.
Local Whisper handles audio transcription locally after its tooling and model are available. Claude still receives frames and transcription through the plugin workflow. Local transcription therefore does not establish privacy for the entire session. Our sandbox did not verify network transfers, model downloads, cloud processing, YouTube acquisition or interactive plugin installation.
What does the audit establish?
Our sandbox recorded 17 npm audit advisories: 9 high, 6 moderate, 2 low and zero critical. Scan coverage is unknown because scannedPackages is absent.
The record provides neither dependency paths nor an exploitability assessment. These are recorded advisories for the tested dependency state, not proof of exploitation or a security assessment of current version 1.4.0. Review and audit the exact dependency tree you intend to use before processing private recordings.
The supplied current issue index lists reports of duplicate MCP server instances after reconnect and empty analysis after an ffmpeg timeout. Those are upstream reports, not behaviors reproduced by our sandbox. Check them during evaluation if reconnects or longer analysis jobs matter to your work.
Which alternatives fit a different job?
Use yt-dlp for media acquisition and Gemini CLI for its enterprise terminal-agent environment. Neither supplied extract establishes an equivalent Claude Code video-review workflow.
yt-dlp is a command-line audio/video downloader. Its README supports choosing it alone when downloading is the task and interpretation happens elsewhere. It also complements Claude Video Vision: the plugin's documented YouTube workflow uses it to acquire media.
Gemini CLI is maintained for enterprise Gemini Code Assist workflows on Google Cloud. Consider that terminal agent when you want its environment and have suitable access. Its supplied README says free individual access and Google AI Pro/Ultra moved to another CLI. It does not verify Gemini CLI as a replacement for this plugin's video workflow.
What do the scores mean?
Setup ease is 2/5 because media tooling and backend configuration remain external. Documentation is 4/5 for its explicit tools, configuration and backend instructions.
The editorial rubric is: 1 means substantial obstacles; 2 means significant qualifications; 3 means mixed supporting evidence; 4 means strong evidence for the dimension; 5 requires unusually complete supporting evidence. These assessments are not performance benchmarks or security ratings.
Community remains 3/5 for visible issue-reporting activity, with adoption, response times and resolution quality unmeasured. Maturity remains 2/5 because the dated test command failed, dependency advisories were recorded, and the interactive workflow remains unverified. Documentation quality does not erase those limitations.
What should you verify next?
Verify ffmpeg and ffprobe are available, configure one audio backend, and rerun tests for your chosen version. Investigate remaining failures individually.
Then inspect a known short clip inside Claude Code. Compare timestamps, frames and transcription with the recording before relying on an answer. This supplies the workflow evidence that our installation and build results cannot provide.