StandRig:AI 直接当 rigger,MCP 原生的 2D 角色建模系统
StandRig: AI as the Rigger — MCP-Native 2D Character Modeling Without the Tedium
📌 开源仓库:StandRig(AI 可编辑 2D 建模核心) GitHub:https://github.com/sayaka-aiart/StandRig 配套编辑器桥接:https://github.com/sayaka-aiart/cubism-api-bridge 作者:sayaka-aiart(开发者预览版 0.2.0,Apache 2.0 / MIT)
BLUF:StandRig 是一个把分层 PSD 转化为可被 AI 直接操作的 2D 角色模型的本地服务。通过 MCP 协议,Claude Code 这类 AI 客户端可以编辑 Mesh、Deformer、Key Form,实时查看结果——不需要人坐在 Live2D 编辑器里逐点拖拽。配套的 cubism-api-bridge 进一步打通了 Live2D Cubism Editor 的官方 External API。
Rigging 的时间消耗去哪了
传统 2D VTuber 模型的 rigging 流程大致是:
- 画师出图,分好图层(几十到上百层)
- Rigger 在 Live2D Cubism 里逐层打点、建 Mesh
- 设置 Deformer(变形器)的父子层级
- 一个参数一个参数地设 Key Form(关键形态)
- 反复测试、细调
整个过程极度精细、高度重复,且需要大量”眼睛看着调”的判断。熟练 rigger 做一个中等复杂度的模型要 40-80 小时。
这正是 AI 最擅长的工作类型:有明确约束、大量重复操作、需要快速迭代验证。
StandRig 的做法
StandRig 把整个流程拆成两部分:建模核心(本地服务) 和 AI 接口(MCP)。
本地服务(http://127.0.0.1:5180):
- 读取分层 PSD,解析图层位置、层级、透明度
- 维护模型状态(Mesh、Deformer、参数、Key Form)
- 提供浏览器 UI 预览和参数滑块
- 管理 Checkpoint 和 Bundle 导出
MCP 接口(AI 的操作端):
- AI 读取
standrig://docs/contract和standrig://docs/guide了解当前模型结构 - 通过 MCP 工具调用执行 Mesh 编辑、Deformer 操作、参数设置
- 支持 dry-run(预演)→ 数值 QA → 确定 → 视觉确认 的迭代流程
- 随时 Restore 到任意 Checkpoint
支持的 AI 变形操作:smooth、relax、inflate、pinch、bend、contour-follow,覆盖了 rigging 过程中最高频的调形操作。
快速上手
前置条件只需要 Node.js 22.12+ 或 24+ 和浏览器,无需 AI 订阅就能先跑样例。
# Clone 后
npm ci
npm run build
npm start
# 然后访问 http://127.0.0.1:5180/
连接 Claude Code 等 MCP 客户端:
{
"mcpServers": {
"standrig": {
"command": "node",
"args": ["/path/to/StandRig/packages/mcp/src/cli.mjs"],
"env": { "STANDRIG_URL": "http://127.0.0.1:5180" }
}
}
}
连上之后,给 AI 的第一句话:
StandRig 的 MCP 资源 standrig://docs/contract 和 standrig://docs/guide 请先读一下,然后用 standrig_context 确认当前模型结构,先不要改任何东西。
cubism-api-bridge:打通 Live2D Cubism Editor
StandRig 本身是独立的建模核心,不直接操作 Cubism Editor 的 .cmo3 文件。但很多创作者的最终输出仍然需要经过 Cubism Editor(导出 .moc3、精细调整物理等)。
cubism-api-bridge 解决了这个问题:
- TypeScript 通用客户端:直接包装 Cubism External API(WebSocket)
- 本地 HTTP 服务器:让 Python SDK 和任意 HTTP 客户端都能操作 Cubism Editor
- StandRig 集成:StandRig MCP 可通过 Bridge 从 Cubism 读取模型信息和临时操作参数
StandRig(MCP服务)
↓ 通过 cubism-api-bridge
Live2D Cubism Editor 5.4(External API)
当前是实验版,针对 Cubism Editor 5.4 alpha2 / External API 1.1.0,实现了 56 个 API 中的 47 个常规操作。Python 使用示例:
npm ci && npm run build # 构建 Bridge
npm run http -- --port 22035 # 启动 HTTP 代理
python -m pip install ./python # 安装 Python SDK
为什么这件事值得注意
这不只是”又一个 VTuber 工具”。从技术路线上看,StandRig 做了一件很重要的事:从第一天就为 Agent 设计接口,而不是给人用的 GUI 加一个 AI 按钮。
模型状态、操作约束、确认/回滚机制——都是面向程序化调用设计的。AI 不是在模拟人点击 Live2D 的界面,而是通过机器可读的 API 直接操作模型数据。
这和我们上篇文章讲的 YC 2026 “Software for Agents” 赛道完全对应:每一类人类在用的创意工具,都需要为 Agent 重做一遍。2D 角色 rigging 这个领域,现在有了第一个认真的答案。
当然,现在还是 0.2.0 开发预览版,有明确的限制:
- PSD 需人工分好图层(自动分层不在范围内)
- 不负责图像生成(只做建模/绑定,不出图)
- Live2D 的
.moc3创建/转换不支持(需要通过 Cubism Editor 完成) - 仅在 Windows / Node.js 24 验证过
这些限制是合理的边界划定,而不是缺陷。工具的核心价值在于把 rigging 这件最耗时的事变成 AI 可以自主完成的任务,并且做到了。
© 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: StandRig — AI-editable 2D modeling core GitHub: https://github.com/sayaka-aiart/StandRig Editor bridge: https://github.com/sayaka-aiart/cubism-api-bridge Author: sayaka-aiart (developer preview 0.2.0, Apache 2.0 / MIT)
BLUF: StandRig is a local service that turns a layered PSD into a 2D character model that an AI can directly operate. Via MCP, Claude Code or any MCP client edits meshes, deformers, and key forms, seeing results immediately — no human dragging points in Live2D Cubism. The companion cubism-api-bridge connects it further to the official Cubism Editor External API.
Where the Rigging Time Goes
Traditional 2D VTuber rigging is roughly:
- Artist delivers art with separated layers (dozens to hundreds)
- Rigger opens Live2D Cubism, places mesh points layer by layer
- Sets up deformer parent-child hierarchies
- Sets key forms parameter by parameter
- Tests, iterates, fine-tunes — visually, by eye
An experienced rigger spends 40–80 hours on a moderately complex model. The work is precise, repetitive, and iterative — exactly the profile where AI performs best.
How StandRig Works
StandRig splits the workflow into two clean parts: a modeling core (local service) and an AI interface (MCP).
Local service (http://127.0.0.1:5180):
- Parses a layered PSD: layer positions, hierarchy, transparency
- Maintains model state: meshes, deformers, parameters, key forms
- Browser UI for preview and parameter sliders
- Checkpoint management and bundle export
MCP interface (what the AI touches):
- AI reads
standrig://docs/contractandstandrig://docs/guideto understand the model - Calls MCP tools to edit meshes, deformers, parameters
- Dry-run → numerical QA → commit → visual verify loop
- Restore to any checkpoint at any time
Supported deformation operations: smooth, relax, inflate, pinch, bend, contour-follow — the highest-frequency shape adjustments in rigging work.
Quick Start
Prerequisites: Node.js 22.12+ or 24+ and a browser. No AI subscription needed to test with the included sample.
npm ci && npm run build && npm start
# Then open http://127.0.0.1:5180/
Connect Claude Code or any MCP client:
{
"mcpServers": {
"standrig": {
"command": "node",
"args": ["/path/to/StandRig/packages/mcp/src/cli.mjs"],
"env": { "STANDRIG_URL": "http://127.0.0.1:5180" }
}
}
}
First message to the AI:
Please read StandRig’s MCP resources standrig://docs/contract and standrig://docs/guide, then use standrig_context to describe the current model’s part structure and what motion is already configured. Don’t change anything yet.
cubism-api-bridge: Connecting to Live2D Cubism Editor
StandRig is a standalone modeling core — it doesn’t directly touch Cubism Editor’s .cmo3 files. But many creators need Cubism Editor for final export (.moc3) and physics tuning.
cubism-api-bridge solves this:
- TypeScript client: wraps Cubism External API (WebSocket) directly
- Local HTTP server: lets Python SDK and any HTTP client operate Cubism Editor
- StandRig integration: StandRig MCP can read model info and temporarily set parameters via the bridge
StandRig (MCP service)
↓ via cubism-api-bridge
Live2D Cubism Editor 5.4 (External API)
Currently experimental, targeting Cubism Editor 5.4 alpha2 / External API 1.1.0, with 47 of 56 APIs implemented.
Why This Matters
This isn’t just another VTuber tool. The design decision that matters: StandRig is agent-first from day one, not a human GUI with an AI button bolted on.
Model state, operation constraints, commit/rollback — all designed for programmatic access. The AI isn’t simulating a human clicking in Live2D’s interface; it’s directly operating model data through machine-readable APIs.
This maps exactly to the YC 2026 “Software for Agents” track: every category of human-facing creative tool needs to be rebuilt for agents. For 2D character rigging, StandRig is a first serious answer.
Current limitations are clearly scoped:
- PSD must be manually layer-separated (auto-segmentation is out of scope)
- No image generation (modeling/binding only)
- No
.moc3creation/conversion (needs Cubism Editor for that) - Tested on Windows / Node.js 24 only
These are deliberate boundaries, not defects. The core value — making rigging, the most time-consuming part, something AI can do autonomously — is delivered.
© 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:hello@mushroom.cv
💬 评论与讨论
使用 GitHub 账号登录后发表评论