VVV（Vibe Vision Video）：模板建一次，跨 agent 反复用。把你的照片和文字换进去就是一部新片；也可以给它一部你喜欢的片子，让它量出来、重建成你自己的版本。

GitHub: https://github.com/actionow-ai/vvv（MIT） ｜ 官网: vibevision.video

中文版

一句话，剩下的 agent 全包

选模板、放照片、写文案、渲染、检查、交付，全在对话里完成。你看样片，说要改什么就行：

用 vvv skill 把 ./photos 里的照片做成一部 9:16 婚礼短片。我们是林一周和徐知夏，10 月 18 日结婚。先给我一个检查过的样片。

安装（任选其一）

• Claude Code：/plugin install vvv --marketplace actionow-ai/vvv（插件自带模板）
• Codex、Cursor、Gemini CLI、Copilot 等 25+ agent：npx skills add actionow-ai/vvv，再 clone 仓库并设置 VVV_KIT
• WorkBuddy：从 Releases 下载 skill ZIP，在「技能 → 添加技能」上传
• 任意 MCP 客户端：claude mcp add --transport http vvv https://mcp.vibevision.video/mcp，免安装浏览模板和指南

渲染在本机跑，需要 Python 3.9+、Node.js 26 和 npm，其余依赖 agent 自己装。

模板只建一次

场景结构、镜头运动、16:9 和 9:16 两种布局、按节拍规划的时序、渲染检查——这些建一次；每部新片只换照片、名字日期、每一句文案、音乐和格式。vvv.py describe 告诉 agent 每个模板的照片槽位、文案字段和设置；validate 在渲染前拦住错误输入。复用的意义就是省下反复重设计、重生成、调同一套视频代码的 token。内置模板包括 Doodle（手绘）、Album（相册）、DV Story（DV 机）、Comedy（喜剧开场）、Starter（卡点混剪）、Album 9:16（竖屏）等，全部通过 vvv check 渲染验证。

给它一部喜欢的片子，还你一个自己的版本

给 agent 一个参考视频加上你的素材（照片、logo、文案、音乐），它研究参考片是怎么做的，重建成可编辑的工程：01 参考 + 素材（参考片只研究、不外发）→ 02 测量（按节拍切段、镜头长度、静帧、节拍对齐，都从真实文件量）→ 03 重建（套最近的模板或新合成，缺的美术素材自动生成，并对照参考片检查）。最好配能生成图片的 agent；Claude Code 和 Codex 都装了的话，Claude Code 可以通过内置桥接用 Codex 的图片生成。视频和语音不生成。

还包含

• 渲染检查：vvv check 测格式、音频、静帧段落和节拍对齐，再列出只有人眼能判断的项
• 交接：模板、特效或成品工程打包成一个 .vvv 文件（含代码、媒体许可和渲染证据），任何 agent 都能接手继续
• 素材工具：本地 FFmpeg 助手，探测媒体、切区间、抽音频、采帧
• 社区模板：自己做的模板打包后在官网提交，审核通过后上架网站和 MCP 服务器

状态

0.1 早期版本。已在 macOS ARM64 上用 Claude Code 和 Codex 端到端验证；WorkBuddy 和其他平台测试较少。代码、文档、模板和演示媒体均为 MIT（字体、预览配乐和第三方代码保留各自条款）。

English Version

VVV (Vibe Vision Video): build a template once, reuse it across agents. Swap in your photos and words for a new film — or bring a film you love and get your own version.

GitHub: https://github.com/actionow-ai/vvv (MIT) ｜ Website: vibevision.video

Say one sentence. Your agent does the rest.

Picking the template, placing photos, writing copy, rendering, checking and delivering all happen in the conversation. You watch the sample and say what to change:

Use the vvv skill to make a 9:16 wedding short from the photos in ./photos. We are Lin Yizhou and Xu Zhixia, married on October 18. Show me a checked sample first.

Add it to your agent (any one of)

• Claude Code: /plugin install vvv --marketplace actionow-ai/vvv (the plugin includes the templates)
• Codex, Cursor, Gemini CLI, Copilot and 20+ more: npx skills add actionow-ai/vvv, then clone this repository and set VVV_KIT
• WorkBuddy: upload the skill ZIP from Releases under Skills → Add Skill
• Any MCP client: claude mcp add --transport http vvv https://mcp.vibevision.video/mcp to browse templates and guides with no install

Rendering runs on your machine, which needs Python 3.9+, Node.js 26 and npm; the agent installs everything else.

One template, every new film

Scene structure and camera moves, layouts for 16:9 and 9:16, beat-planned timing and render checks are built once. A new film changes only photos, names and dates, every line of copy, music and format. vvv.py describe tells the agent each template’s photo slots, copy fields and settings; validate catches bad inputs before a render. Reuse saves the tokens otherwise spent redesigning, regenerating and debugging the same video code. Built-in templates include Doodle (hand-drawn), Album, DV Story (camcorder), Comedy (comic opener), Starter (beat-cut reel) and Album 9:16 (vertical short), all rendered straight from the templates and passing vvv check.

Show it a film you love. Get your own.

Give your agent a reference video plus your attachments (photos, logo, copy, music). It studies how the reference works and rebuilds it as an editable project: 01 Reference + attachments (the reference is studied, never shipped) → 02 Measure (sections in beats, shot lengths, still frames and beat alignment, measured from the actual file) → 03 Rebuild (on the closest template or a new composition; missing artwork is generated; the result is checked against the reference). Best with an image-capable agent — with Claude Code and Codex both installed, Claude Code can use Codex’s image generation through the bundled bridge. Video and voice are not generated.

Also included

• Render checks: vvv check measures format, audio, still stretches and beat alignment, then lists what only a viewer can judge
• Handoff: pack a template, effect or finished project as one .vvv file with its code, media licenses and render evidence, so any agent can continue it
• Footage tools: a local FFmpeg helper probes media, cuts ranges, extracts audio and samples frames
• Community templates: pack yours and submit it on the website; once approved, it appears on the site and in the MCP server

Status

0.1, an early release. Verified end to end on macOS ARM64 with Claude Code and Codex; WorkBuddy and other platforms are less tested. Code, docs, templates and demo media are MIT; fonts, preview soundtracks and third-party code keep their own terms.

来源 / Source

• GitHub: actionow-ai/vvv（MIT）
• 官网 / Website: vibevision.video