
Model Lab
io.github.hpdkhoa由 GitHub 账号 hpdkhoa 发布v0.2.0更新于 Oct 24, 2026
Draws rigged game gun models in Blender from parameters, checked against the real gun
概览
按参数在 Blender 中生成带骨骼绑定的 AR-15 系列游戏枪械模型,并自带自检与实枪照片轮廓比对。
- 可以做什么
- 列出预设枪型、枪族、参数说明以及导轨与 M-LOK 附件类别
- 按预设或枪族构建枪械,可修改参数,并可选生成动作 GIF
- 生成经过测量的动画片段(拉栓、射击、换弹、枪托位置),输出 FBX 与 GIF
- 将构建结果的轮廓与内置实枪参考照片进行比对
- 列出历史构建并读取其事实数据、标记点与检查结果
- 适用场景
- 适合需要带骨骼、附件插槽和动画片段的占位或原型游戏武器素材,并希望几何形状对照公开尺寸与参考照片进行校验的场景。面向制作 AR-15 类步枪的游戏美术与开发者,不适用于其他枪族或制造数据需求。
- 运行要求
- 以 Python 包(modellab-mcp)通过 uvx 或 pip 在本地以 stdio 运行。需要 Python 3.10 及以上和 Blender 4.2 及以上。可选环境变量:MODELLAB_BLENDER 指定 Blender 可执行文件路径,MODELLAB_OUT 指定唯一输出目录(默认 ~/.modellab),MODELLAB_WAIT_FOR 在指定进程运行时阻止构建。仅支持桌面端。
安装
在 SourceWeft 中
- 打开 控制台中的 Model Lab,将其添加到工作区。
- 为需要使用其工具的对话启用该服务。
Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。
其他 MCP 客户端
参照 仓库 中的启动说明。
README
Model Lab
Game gun models drawn in Blender from parameters, rigged and checked against the real gun.
[license: MIT] [Python 3.10+] [PyPI: modellab-mcp] [MCP Registry: modellab-mcp] [Blender: 4.2+] [tools: 10] [presets: 5 AR-15s]
An MCP server that draws game gun models in Blender from a few parameters. Each build has real parts at real proportions, one bone a moving part, its clips, and the facts a game needs. Every build checks itself, and a second check measures its outline against a photo of the real gun.
[An AR-15 with a holographic sight, a magnifier, a light and a vertical grip]
Ask for "an M4A1 with a red dot and a light on the right rail" and Model Lab writes a rigged FBX, an FBX for each
attachment, a .blend, a sheet of pictures and a facts.json with every mark in gun space. It tells you which checks
failed, if any. It makes game models only: outside shapes and moving parts. It draws no working internals and
writes no fabrication data.
What a build gives you
Gun space: centimetres, +X the gun's left, +Y forward, +Z up. The origin is the bolt face at rest on the bore.
Install
Model Lab needs Python 3.10 or later and Blender 4.2 or later.
or
Add it to your MCP client (Claude Code, Claude Desktop, or any client that runs stdio servers):
The tools
The same calls run from a shell: modellab presets, modellab build M4A1 stock_pos=2 name=MyM4,
modellab clips <build>, modellab compare. Run modellab help for the list.
Every build checks itself
A build reports each check with its value. The checks:
- the length against the published figure (the M4A1 at 29.75 and 33 in, the M16A2 at 39.63 in);
- the chamber on the bore;
- the dust cover: closed over the whole port, hinged on its bottom edge, and open against the lower, not through it (the build swings it a degree at a time and stops it where it first touches);
- the magazine clear of its well's walls and seated;
- the bolt's travel past the top round;
- the sight line: the rear aperture and the front post at one height;
- each attachment on a rail it fits, clear of the gun in every state, and its optic on the iron sight line;
- every visible part with its real features (
parts_complete). The parts table lists each part a person sees on the real gun and the features that make it read as that part: the charging handle's swept wings, finger hooks, latch and roll pin; the bolt catch's serrations; the castle nut's notches; the front sight base's bayonet lug, swivel and taper pins; and so on. A missing feature fails the build. Marks such as rail numbers are listed as texture, not shape.
[Close views: the charging handle]
The real-gun check
compare_to_reference draws the build's side outline at a reference photo's own scale. It lines the build up on the
muzzle and the bore, measures both outlines column by column, and names the parts where they differ. The overlay
shows the model in see-through red over the photo: dark photo showing through is real gun the model misses.
[The M4A1's outline over the reference photo]
This check found the faults the self-checks could not: the lower 1.55 cm too far back, the front sight base 2.4 cm back and the wrong shape, the stock's outline, the magazine's curve, and the rail's height. Now no stretch of the gun's outline over 1.75 cm long is more than 1 cm off. Attachments are classes, not the photo's exact models, so the check keeps their outline apart.
The reference ships with the package: a photo of an M4A1 carbine from PEO Soldier, File:M4A1-flattop.jpg on Wikimedia Commons, in the public domain as a work of the US Federal Government. The check reads its scale from the rail's teeth, 0.394 in apart.
Attachments on sockets
Each rail reports its slots at the Picatinny pitch (MIL-STD-1913), slot 0 at the rear. M-LOK handguards report their
slots too (32 by 7 mm, 40 mm apart). Pass attachments by name, or as {item, rail, slot, state, options}:
The nine classes: folding rear and front sights, the M4's carry handle, a 30 mm red dot with flip caps, a holographic sight, a 3x magnifier on a flip mount, a 4x32 prism sight, a vertical grip, a weapon light. A build checks that each fits its rail, clears the gun in every state, and clears the others as shown. When one blocks another's move, the build lists it: behind a magnifier, the rear sight goes up only after the magnifier swings aside.
[The sight picture through a red dot, the A2 post in its middle]
Motion and clips
Every .blend holds the cycle as an action. motion=true renders it as GIFs: the charging handle drawn back, the
cover springing open, the carrier flying home, the magazine out and in, then the attachments' own moves.
[The cycle on an AR-15 with a magnifier and a holographic sight]
build_gun_clips keys the gun's clips at 60 frames a second from real timings: charge, chamber check, fire (a 75 ms
cycle), the last shot (the bolt held on its catch), reload, empty reload, and the stock's positions. Each clip is
measured: every part on its own axis, the full stroke, the cover open before the carrier has moved 1 cm, the magazine
out of the well and seated, the bolt held until the magazine is in.
[An empty reload: the bolt held open until the new magazine is seated]
Next: soft edges where a hand holds
The owner, 2026-10-21: "the grip is no sharp edge, this will be the next improvement for Model Lab when read on the manual or images, make the corner round a bit, especially the places that human grab and hold, most of the time they will no sharp edge because sharp edges will hurt the hand/human, make them softer so hand can hold it".
- Every part a hand grabs or holds gets rounded edges: the grip, the pistol grip's panels, the handguard, the magazine, the charging handle, the stock's cheek and butt.
- Each radius is read from the manual, drawing or photo when it shows one. Otherwise it comes from a named human-engineering source, recorded with the part.
- The real-gun check compares these edges too.
Why it matters, measured: on a digitized M1911A1 the grip safety was a square extrusion. A hand placed on the grip had the safety's edge 0.42 cm in its palm (the palm gives 0.27 cm). Rounding the grip parts by 0.75 mm, the least MIL-STD-1472G 5.7.7.6 allows, cut that to 0.09 cm past the give. Red: before (sharp). Blue: after (rounded). Below, close-ups of the spot before and after.
[The palm's clip against the grip safety, sharp and rounded]
Limits
- One platform so far: the AR-15 family. Others will follow.
- One reference photo: the M4A1 from its right side.
- The clips are the gun's side only. A hand's clips need an arms rig.
- No selector clip yet: the lever's direction at SAFE is not yet checked against a source.
- Many numbers are estimates sized from photographs.
describe_paramslists every source and every estimate.
Trademarks
M-LOK is a trademark of Magpul Industries. Model Lab draws M-LOK slots to their published dimensions and is not connected with Magpul. AR-15 is a trademark of Colt's Manufacturing; here it names the rifle family in its common sense. Military designations (M4A1, M16A2, M68) name the guns and parts Model Lab copies. No product is endorsed.
License
MIT. See LICENSE.
来源:README.md,提交 b5098b8
工具
0版本历史
2- v0.2.0最新Oct 15, 2026
- v0.1.0Oct 14, 2026
评分
已满足适用检查项的 50%
该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。
- 近 30 天内可连通25 分
请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。
- 已声明认证方式10 分
请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。
- 已列出工具15 分
请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。
- 描述必需15 分
- 图标10 分
- 文档15 分
- 许可证10 分
其他 Model 集成

model.tools — Zero-Auth Agent Utilities
tools.model官方发布者
零认证远程工具服务器,为智能体提供一次性邮箱验证码读取、持久检查点、Webhook 收件箱、定时唤醒、任务回执与网页转 Markdown。
NLP.COM.AI Open Model Catalog
ai.com.nlp官方发布者
搜索精选的开放权重AI模型发布、许可证、架构和原始来源。AI 翻译

Hologram Model Hub
foundation.uor.hub官方发布者
一个远程 MCP 服务器,让助手搜索开放模型索引,并获取带 SHA-256 校验值的下载链接。

ModelLens
由 GitHub 账号 Greatbeing 发布
探测兼容 OpenAI 的模型网关,按实测延迟与成功率给模型排名,并推荐该用哪个模型。

Bedrock Multi-Model
由 GitHub 账号 Sagargupta16 发布
本地 MCP 服务器,把 AWS Bedrock 的文本、图像、视频和嵌入模型封装成九个工具,让助手可以提问、对比和评审模型。