Hypit：给 Claude Code 和 Codex 一套视频制作语言，一条命令复刻爆款
Hypit: A Video Workflow Language for Claude Code and Codex — Clone Any Viral Video in One Command
📌 开源仓库：hypit-ai/hypit GitHub：https://github.com/hypit-ai/hypit 官网：https://hypit.ai License：Apache-2.0 with conditions | Stars：3860 | Language：TypeScript
Hypit 要解决的问题：AI 编程 Agent 没有视频制作的语言。
文字 Agent 有 Markdown。代码 Agent 有编程语言。但视频——拼接时间线、对齐字幕、替换 B-roll——一直是手工操作或脚本碎片，不是 Agent 能直接读写的结构化表达。
Hypit 给 Claude Code、Codex 这类 Coding Agent 提供了 SVML（视频工作流语言）和一套渲染系统。
核心概念：词语锚定，不是秒数锚定
传统视频编辑的时间线以「秒」为单位——剪辑点在第 3.5 秒，字幕出现在第 5.2 秒。这对代码生成来说非常难处理：语言模型写出来的时间数字几乎没有意义。
Hypit 的 SVML 把视频元素锚定在词语上。字幕跟着台词走，B-roll 出现在关键词那句话期间，音效在某个词被说出来时触发。修改脚本之后，整个时间线自动重新对齐——不用手动调每个剪辑点。
左侧 SVML 源文件 → 右侧实时渲染视频Agent 写 SVML，Hypit 渲染出视频。
安装和使用
npx skills add hypit-ai/hypit -g这条命令把 Hypit 安装为 Coding Agent 的 Skill。首次使用时，Agent 会检查 Hypit 可执行程序并按需安装。
安装后，在任意项目目录对 Agent 说：
/hypit Clone this video: /path/to/videoAgent 分析视频，提取 workflow（画面结构、字幕风格、B-roll 逻辑、特效），输出可编辑的 SVML 文件。也可以不传参考视频，直接描述：
/hypit Make a ranking video that puts Hypit in S tier.Agent 从零写 SVML，Hypit 渲染成视频。
Hypit 本身免费，模型服务（Seedance、GPT Image 等）按各自定价计费。官方推荐 HypiHub 托管服务，也可以配置自己的 API 或本地模型。
三个实测例子（成本来自官方 README）
GOAT DEBATE（足球排行榜，$1.15） 20 秒，把 Ronaldo 排进 D 级、Messi 排进 S 级。A-roll 是两段 Seedance 2 Mini 720p 生成的片段，B-roll 是 GPT Image 2 生成的 2K 肖像和 10 张 1K 脑腐图，WhisperX 逐词对齐，64 个无头 Chromium 并发渲染。三个变体：换解说员、翻转排名、换科技公司创始人。
DAILY CREATINE（播客片段，$1.07） 18 秒，分屏采访布局，逐词字幕。三个变体：Pepe 和 Doge 争狗腿力量、换成护肤品广告、换成 CheatGPT 应用广告。
NICE RIDE（街头采访，$1.09） 26 秒，头部追踪字幕（调用 Google Video Intelligence 和 YOLOv8 识别人脸坐标），三个变体包括西班牙语版本——台词重写后时间线自动重新对齐。
支持的场景
| 场景 | 说明 |
|---|---|
| 付费社交广告 | 从 Meta 广告库克隆一条高转化广告，换产品，同天出 50 个 hook 变体 |
| 爆款复刻 | TikTok、Reels、Shorts，换主播/话题/语言/比例，结构不变 |
| TikTok Shop 带货视频 | 一套格式，每天换 SKU，CTA 结构保持 |
| AI UGC / 讲解视频 | 旁白、字幕、B-roll、评论贴纸、节拍切割，自动串联 |
| 播客和采访片段 | 分屏布局，分说话人字幕，反应叠加 |
| 纯代码渲染视频 | 不调任何生成 API，只用前端代码渲染画面 |
| 本地化版本 | 同一条视频出十种语言，改一行脚本时间线自动重排 |
技术架构（从 README 拆解）
- 语言：TypeScript 5.9，Node.js 22.15+，pnpm 10.33
- 渲染：64 个无头 Chromium 进程并发渲染
- 字幕对齐：WhisperX 逐词对齐（非 segment 级别）
- 人脸检测：Google Video Intelligence + YOLOv8 AnimeFace（用于跟踪字幕位置）
- 接入方式：
npx skills add安装为 Skill，Agent 通过/hypit调用 - 模型接口：可插拔，支持 HypiHub、自有 API 或本地模型
许可证拆解（Apache-2.0 with conditions）
“Apache-2.0 with conditions”不等于完全自由。主要限制条款：
允许（无需商业授权）：
- 在自己的基础设施上运行，做自己组织的商业项目
- 为客户生产视频（单租户）
- 修改源码并在相同 License 下开源
需要商业授权：
- 多租户 SaaS：如果你把 Hypit 包装成平台，让两个及以上独立组织使用各自的 workspace，属于多租户服务，须获得 Hypit.AI 书面授权。是否收费无关紧要。
- 商业再分发：把 Hypit 打包销售或作为收费产品的一部分，须获得授权。
关于生成内容的归属：协议明确说明，产出的视频、音频、图像全部归创作者所有，Hypit.AI 不主张任何权利。
适合谁
- 需要批量出视频变体的内容团队
- 想用 Claude Code / Codex 直接操作视频 workflow 的开发者
- 需要把 AI 生成素材（图像、语音）组装成完整视频的工程师
生成视频的部分不是 Hypit 的核心——核心是让 Agent 能写、读、修改视频 workflow。把生成这步拿掉，纯用代码渲染的视频也能跑通全流程。
开源代码与模型仅供学习、勿直接用于工作。
© 2026 Author: Mycelium Protocol. 本文采用 CC BY 4.0 授权——欢迎转载和引用，须注明作者姓名及原文链接，不得去除署名后以原创发布。
📌 Repository: hypit-ai/hypit GitHub: https://github.com/hypit-ai/hypit Website: https://hypit.ai License: Apache-2.0 with conditions | Stars: 3,860 | Language: TypeScript
Hypit solves a gap: AI coding agents have no language for making videos.
Text agents have Markdown. Code agents have programming languages. But video — splicing timelines, aligning captions, swapping B-roll — has always been manual work or fragmented scripts, not a structured representation that an agent can read and write.
Hypit gives AI coding agents (Claude Code, Codex) SVML — a video workflow language — plus a rendering system.
Core Concept: Word-Anchored, Not Second-Anchored
Traditional video timelines use seconds as their unit — a cut at 3.5s, a caption appearing at 5.2s. This is nearly unusable for code generation: timestamps generated by a language model are almost always wrong.
SVML anchors video elements to words. Captions follow dialogue. B-roll appears during the line containing the relevant keyword. Sound effects fire when a specific word is spoken. When you rewrite the script, the entire timeline realigns automatically — no manual adjustment of every cut.
SVML source (left) → Live video render (right)The agent writes SVML; Hypit renders the video.
Install and Use
npx skills add hypit-ai/hypit -gThis installs Hypit as a Skill for your coding agent. On first use, the agent checks for the Hypit executable and helps install it if needed.
After installation, tell your agent from any project directory:
/hypit Clone this video: /path/to/videoThe agent analyzes the video, extracts the workflow (scene structure, caption style, B-roll logic, effects), and outputs an editable SVML file. Or skip the reference video and just describe:
/hypit Make a ranking video that puts Hypit in S tier.The agent writes SVML from scratch; Hypit renders it to video.
Hypit itself is free. Model services (Seedance, GPT Image, etc.) are billed by the service you choose. The official recommendation is HypiHub (managed service), but you can configure your own API or local models.
Three Real Examples (Costs From the Official README)
GOAT DEBATE (Football tier list, $1.15) 20 seconds. Ronaldo in D, Messi in S. Two A-rolls from Seedance 2 Mini 720p, one 2K portrait and ten 1K brain-rot B-rolls from GPT Image 2, WhisperX word-level alignment, 64 concurrent headless Chromium processes. Three clones: swap narrator, flip the rankings, swap players for tech founders.
DAILY CREATINE (Podcast clip, $1.07) 18 seconds. Split-screen interview layout, speaker-aware captions. Three clones: Pepe and Doge argue about forearms, retinol skin-care ad, CheatGPT app demo.
NICE RIDE (Street interview, $1.09) 26 seconds. Head-tracked captions (Google Video Intelligence + YOLOv8 face bounding boxes). Three clones including a Spanish version — rewrite the lines and the timeline realigns itself.
Supported Scenarios
| Scenario | Description |
|---|---|
| Paid social ads | Clone a high-converting ad from Meta Ad Library, swap product, ship 50 hook variants same day |
| Viral clones | TikTok, Reels, Shorts — swap host/topic/language/aspect ratio, keep the structure |
| TikTok Shop affiliate videos | One format, one new SKU per day, the conversion structure unchanged |
| AI UGC / talking head | Narration, captions, B-roll, comment stickers, beat-synced cuts, auto-wired |
| Podcast and interview clips | Split-screen layouts, per-speaker captions, reaction overlays |
| Code-rendered videos | No generation API calls at all — pure front-end code renders the visuals |
| Localized versions | Same video in ten languages; edit one line and the timeline re-flows |
Technical Architecture (From the README)
- Language: TypeScript 5.9, Node.js 22.15+, pnpm 10.33
- Rendering: 64 concurrent headless Chromium processes
- Caption alignment: WhisperX word-level alignment (not segment-level)
- Face detection: Google Video Intelligence + YOLOv8 AnimeFace (for tracking caption position)
- Integration: Installed as a Skill via
npx skills add, invoked via/hypit - Model interface: Pluggable — HypiHub, your own API, or local models
License Breakdown (Apache-2.0 with Conditions)
“Apache-2.0 with conditions” is not the same as fully free. The key restrictions:
Allowed without a commercial license:
- Run on your own infrastructure for your own organization’s commercial projects
- Produce videos for clients (single-tenant)
- Modify and republish source code under the same license
Requires a commercial license:
- Multi-tenant SaaS: If you wrap Hypit into a platform where two or more independent organizations share separate workspaces, that’s a multi-tenant service and requires written authorization from Hypit.AI — whether or not you charge a fee.
- Commercial redistribution: Selling Hypit or packaging it as part of a paid product requires authorization.
On output ownership: The license explicitly states that all videos, audio, images, and other artifacts you produce with Hypit belong to you. Hypit.AI claims no rights in your content.
Who It’s For
- Content teams that need to produce video variants at scale
- Developers who want Claude Code or Codex to directly write and edit video workflows
- Engineers assembling AI-generated assets (images, audio) into finished videos
Video generation isn’t Hypit’s core — the core is giving agents a structured language to write, read, and modify video workflows. Remove the generation step entirely and you can still run the full pipeline with code-rendered visuals only.
Open-source code and models are for learning purposes only — do not use directly in production work.
© 2026 Author: Mycelium Protocol. Licensed under CC BY 4.0 — free to share and adapt with attribution. You must credit the author and link to the original; removing attribution and republishing as original is not permitted.
关于本站 · 免责声明
🍄 Mushroom Research Blog 是非营利、免费公开的个人科技观察博客与公众号 XStack18，不接受商业合作、不代表任何企业或机构立场，也不谋求商业利益。我们以个人视角客观中立地记录和分析 AI、Web3 等领域的最新模型发布与技术动态——不止转述新闻标题或二手信息，而是给出有独立思考的深入分析，希望帮更多人获得有价值的一手科技认知。
⚠️ 文中介绍的开源代码与模型，仅供学习交流与技术借鉴。它们大多仍处于早期阶段，有待进一步研究和验证，请勿直接用于工作或生产环境；如需采用，请先自行充分测试，并核实其许可证与安全性。
Open-source code and models featured here are shared for learning and reference only. Most are early-stage and still need further study and verification — please don't use them directly in your work or in production. Test them thoroughly and check their licenses and security first.
- 本站文章均为作者基于公开信息的个人研究与观点整理，不代表文中提及的任何公司、产品、模型的官方立场，未与其构成商业关联或合作关系。
- 科技行业信息更新极快，我们尽力保证内容准确、及时，但不对完整性、实时性做绝对保证，具体请以相关企业/项目官方公告为准。
- 文中引用的第三方商标、产品名称、图片、数据等版权归原权利人所有，我们会尽量注明来源；如你认为存在版权疑问或侵权，请通过下方邮箱联系我们，收到通知后会尽快核实处理（更正、加注来源或删除）。
- 文章内容仅为技术科普与个人观点，不构成投资、法律或其他专业建议，据此进行任何决策的后果需自行判断和承担。
📮 侵权 / 勘误 / 合作咨询：[email protected]
💬 评论与讨论
使用 GitHub 账号登录后发表评论