Jev directory
CommunityJev Agents & MCP
Claude Video Review Skill
Claude Code skill that reviews a finished video before publishing: Jev checks every frame against character reference sheets, Gemini watches with sound, GPT reviews still frames, and Claude returns a timestamped repair list.
- Stars
- 0
- Reading time
- 2 min
- Published by
- Community
- Added
- Nov 2026
Highlights
- Four layers: measurement scripts, Jev frame-by-frame checks, blind Gemini and GPT reviews, and a Claude confirm-and-decide step.
- Jev checks every frame for wrong faces, extra people, silent-shot mouth movement, broken hands, garbled text, and missing key actions.
- A 30-second video at 24 fps has 720 frames, and most generation glitches last one to four frames, so no sampling is used.
- Runs locally with your own keys: GEMINI_API_KEY required, TYPESAFE_API_KEY recommended, GPT via your own Codex CLI sign-in.
- Requires Claude Code, Python 3.10+, and ffmpeg; released under the MIT licence by Poorna Reddy at Arkya AI.
Questions about this entry
Short answers drawn from the linked project or page.
It reviews a finished video before publishing using three AI reviewers: Jev by TypeSafe AI checks every frame, Gemini watches the video with its sound, and GPT reviews still frames. Claude then returns a verdict plus a list of problems with timestamps and a fix for each.
Clone the repo into ~/.claude/skills/video-review, run pip install -r requirements.txt, then copy .env.example to .env and fill in your keys. After that you can ask Claude Code to review a file such as final.mp4.
Gemini is required and uses GEMINI_API_KEY; Jev by TypeSafe AI is recommended and uses TYPESAFE_API_KEY in the same .env file; GPT review uses your own OpenAI sign-in through codex login. The skill ships with no keys.
A 30-second video at 24 fps contains 720 frames, and most AI generation glitches last only one to four frames, so the skill checks every frame rather than a sample.
Scripts using ffmpeg, Python, OpenCV, and Whisper check loudness against -14 LUFS, true peak, clicks, silences, black and frozen frames, duplicated frames (judder), jumps at clip seams, and the transcript against the script.
Claude issues SHIP, FIX, REJECT, or INCONCLUSIVE. It confirms a problem when a measurement and a reviewer agree, when two reviews agree, or when the frames themselves show it.
The skill needs Claude Code, Python 3.10 or newer, and ffmpeg. It is MIT licensed and was built by Poorna Reddy at Arkya AI, with keys kept out of version control via .gitignore.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Vercel's fx safety reviewer, measured at 18× faster
Reddit: Early experiments wiring Jev into Pi as a safety layer and model router.
YouTube: Ray Amjad puts Jev inside an agentic coding loop and looks at what it costs to run.
YouTube: RoboNuggets puts Jev to work inside Claude Code.
More Jev resources like this
Claude Code plugin that puts Jev behind six hooks: request triage, automatic effort, a destructive-command guard, rule review, and a final completion check. One file, zero dependencies, MIT licensed.
Jev Agents & MCP#claude-code#guardrails#coding-agent
CommunityCivalski/claujev
OpenCode workflow harness that adds a durable Markdown work queue, risk-scaled review, six portable skills, four local plugins, and Jev as an advisory typed-question reviewer for diffs, plans, and close states.
Jev Agents & MCP#github#agents#coding-agent
Communityezcodelab/libreharness
An MCP server that exposes TypeSafe's Jev model as twelve judgment tools, including claim verification, injection screening, reranking, classification, extraction, and patch review, each returning typed probabilities in roughly 150 to 500 ms.
Jev Agents & MCP#mcp#agents#typescript
Community
Reusable project context and specialist workflows for the AI tools you already use: agents, skills, rules, and MCP tools for Claude Code, Codex, Copilot, and Cursor, plus Jev decision support that offloads bounded LLM decisions to save tokens.
Jev Agents & MCP#community#github#agents
Community
jmagly/aiwgTypeScript
jmagly/aiwgTypeScriptGitHub issue triage agent built with Eve and Jev. It takes one typed decision per issue through a pipeline of classify, validate, duplicate and apply steps, with dry-run defaults, a label allow-list, and raw Jev answers stored for threshold tuning.
Jev Agents & MCP#agents#github#madewithjev
Communitybenjamincanac/tia
Claude Code and Codex skill that audits a codebase for LLM call sites, proposes which routing, classification, and rubric-scoring decisions to convert to Jev, and reports latency and cost comparisons after the scope is approved.
Jev Agents & MCP#github#claude-code#codex
CommunityKevin-Zhouu/jevify
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.