vrm-toolkit-mcp
An MCP server that enables AI agents to inspect, validate, and preview VRM/VRMA 3D avatar files locally, with tools for metadata extraction, validation, and rendering snapshots or animations via headless Chrome.
README
vrm-toolkit-mcp
An MCP (Model Context Protocol) server that lets an AI agent inspect, validate and preview VRM / VRMA files on your machine. Five tools, stdio transport, no network access at runtime.
VRM / VRMAファイルの検品・情報抽出・プレビューをAIエージェントから使えるようにする MCPサーバーです。ツールは5本、stdioトランスポート、実行時のネットワークアクセスなし。
Tools / ツール
| Tool | What it does | 説明 |
|---|---|---|
vrm_inspect |
Structured JSON: meta (name/author/license and permission fields), humanoid bone map, expressions, polygon/texture statistics | メタ(名前/作者/ライセンス許諾)・ボーン・表情・ポリゴン/テクスチャ規模を構造化出力 |
vrm_validate |
Finding list: missing required humanoid bones, broken node references, missing meta/license fields, unknown required extensions, configurable size/triangle/texture budgets | humanoidマッピング欠落・参照切れ・メタ/ライセンス欠落・未知の必須拡張・サイズ超過などの検品リスト |
vrma_inspect |
Tracks, targeted humanoid bones, expression tracks, duration, estimated frame rate | トラック・対象ボーン・表情トラック・尺・フレームレート推定 |
vrma_preview |
Applies a VRMA to a VRM and renders a PNG still or a short WebM clip (max 10 s) through local headless Chrome/Edge | VRMAをVRMに適用し、ローカルのヘッドレスChrome/EdgeでPNG静止画または短尺WebM動画を書き出し |
vrm_snapshot |
Renders a rest-pose PNG with optional expression weights and view presets (front / three-quarter / side / face close-up) | 表情ウェイトとビュープリセット(正面/斜め/側面/顔アップ)を指定できる静止画スクリーンショット |
Structural tools (vrm_inspect, vrm_validate, vrma_inspect) parse the GLB/glTF container
directly in Node and need no browser. Rendering tools (vrma_preview, vrm_snapshot) use an
already-installed Chromium-based browser via puppeteer-core; nothing is downloaded.
構造系3ツールはNode単体でGLB/glTFを直接パースし、ブラウザ不要。描画系2ツールは
インストール済みのChromium系ブラウザをpuppeteer-coreで使います(追加ダウンロードなし)。
Supported inputs: VRM 0.x (extensions.VRM), VRM 1.0 (extensions.VRMC_vrm),
VRMA (extensions.VRMC_vrm_animation), GLB container or plain glTF JSON.
Requirements / 動作条件
- Node.js 20+
- For
vrma_preview/vrm_snapshot: a local Chrome or Edge (auto-detected; override with theVRM_TOOLKIT_CHROMEenvironment variable) - Input size budget: 512 MiB by default (
VRM_TOOLKIT_MAX_BYTESto override)
Setup / セットアップ
Register with Claude Code / Claude Codeへの登録:
claude mcp add vrm-toolkit -- npx -y vrm-toolkit-mcp
Other MCP hosts / その他のMCPホスト: command npx, args ["-y", "vrm-toolkit-mcp"] (stdio).
From source / ソースから:
npm install
npm run build
claude mcp add vrm-toolkit -- node /absolute/path/to/vrm-toolkit/dist/main.js
Execution examples / 実行例
Real outputs from the local test run (2026-08-05, file paths shortened). ローカルテスト実行(2026-08-05)の実出力です(パスは短縮表記)。
vrm_inspect
Input: a VRoid Studio-exported VRM 1.0 avatar (15.7 MB).
{
"file": { "path": "...\\V1.vrm", "bytes": 15737992, "container": "glb" },
"vrm": { "specFamily": "vrm1", "specVersion": "1.0" },
"meta": {
"name": "V1",
"authors": ["D"],
"licenseUrl": "https://vrm.dev/licenses/1.0/",
"avatarPermission": "onlyAuthor",
"commercialUsage": "corporation",
"creditNotation": "required",
"allowRedistribution": false,
"modification": "prohibited"
},
"humanoid": {
"mappedBoneCount": 54,
"bones": [
{ "bone": "hips", "node": 1, "nodeName": "J_Bip_C_Hips" },
{ "bone": "spine", "node": 34, "nodeName": "J_Bip_C_Spine" }
],
"missingRequiredBones": []
},
"expressions": { "presetCount": 14, "presets": ["happy", "angry", "sad", "..."] },
"geometry": {
"meshCount": 3,
"totalVertices": 103489,
"totalTriangles": 33484,
"perMesh": [
{ "name": "Face (merged)", "vertices": 33640, "triangles": 7308 },
{ "name": "Body (merged)", "vertices": 57128, "triangles": 11548 },
{ "name": "Hair001 (merged)", "vertices": 12721, "triangles": 14628 }
]
},
"textures": { "imageCount": 26, "totalImageBytes": 8033934, "maxDimension": 2048 },
"materials": { "materialCount": 17 }
}
(Output truncated for readability; the tool returns the full bone and image lists.)
vrma_inspect
{
"file": { "path": "...\\motion-walk-forward.vrma", "bytes": 3268, "container": "glb" },
"vrma": { "specVersion": "1.0" },
"humanoid": { "mappedBoneCount": 16, "bones": ["hips", "spine", "chest", "head", "..."] },
"animation": {
"animationCount": 1,
"channelCount": 2,
"channelsByPath": { "translation": 1, "rotation": 1, "scale": 0, "weights": 0 },
"totalKeyframes": 62,
"durationSec": 1.25,
"fpsEstimate": 24,
"frameCountEstimate": 31,
"humanoidBoneTracks": [
{ "bone": "hips", "paths": ["translation"] },
{ "bone": "chest", "paths": ["rotation"] }
]
}
}
vrm_validate
A finding list with stable codes and severities:
{
"ok": false,
"errorCount": 1,
"warningCount": 1,
"findings": [
{ "severity": "error", "code": "HUMANOID_REQUIRED_BONE_MISSING",
"message": "required humanoid bone is not mapped: leftHand (vrm1 required set)" },
{ "severity": "warning", "code": "META_LICENSE_OTHER_URL_MISSING",
"message": "licenseName is Other but otherLicenseUrl is empty" }
]
}
vrm_snapshot / vrma_preview
{
"output_path": "...\\v1-three-quarter-snapshot.png",
"format": "png", "bytes": 194005, "width": 768, "height": 1024,
"view": "three_quarter", "pose": "rest", "humanoid_bone_count": 54
}
{
"output_path": "...\\v1-walk-preview.webm",
"format": "webm", "bytes": 313597, "width": 768, "height": 768,
"clip_duration_sec": 1.25, "recorded_duration_sec": 3, "fps": 30
}
Error codes / エラーコード
Every failure is a structured tool error with one of:
FILE_NOT_FOUND, SIZE_BUDGET_EXCEEDED, NOT_A_GLB, NOT_A_VRM, NOT_A_VRMA,
BROWSER_NOT_FOUND, RENDER_FAILED, OUTPUT_WRITE_FAILED.
The server process never crashes on a bad input file.
Privacy / プライバシー
All processing is local. The server opens no network connection, sends no telemetry, and writes only the output files you name. / 処理はすべてローカルです。ネットワーク接続・テレメトリ送信は なく、書き込みは指定した出力ファイルのみです。
Development / 開発
npm test # build + 25 self-verification tests (structural, boundary, headless rendering)
Rendering tests skip with a notice when no local Chrome/Edge is found.
License / ライセンス
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。