MCP Image Gen 268bb901

MCP设计创作

提供方 aitherlab-dev

设计创作 未标注 无需鉴权 社区 / 第三方 Claude DesktopClaude

一个基于stable-diffusion.cpp和diffusion-rs的本地图像生成服务,无需API密钥和云端支持,所有数据均在本地处理。

一句话选型结论

按资料字段整理

MCP Image Gen 268bb901 属于设计创作类 MCP 服务器,传输方式未标注,无需鉴权,可直接接入 Claude Desktop、Claude,没有现成安装命令,需按官方文档接入。接入前先确认权限范围;涉及写入、删除或付款的操作,建议先在测试环境验证。

分类设计创作 传输方式未标注 鉴权无需鉴权 部署方式未标注
GitHub 星标 1 仓库热度,无公开仓库时显示未标注
工具数 3 暴露给模型的工具数量
资源 / 提示词 0 / 0 资源与提示词模板数量
维护方 社区 / 第三方 核验等级 仅资料核对

工具与能力

设计创作
数量来自来源目录的标注,具体工具清单以官方文档为准
工具数量3暴露给模型可调用的工具个数
资源数量0可读取的资源条目
提示词数量0内置提示词模板数量

能力标签

支持客户端

适用场景

图像生成Rust本地部署人工智能本地处理AI助手工具模型管理

安装与配置

未标注 无需鉴权
上游未给出传输方式
  1. 1
    准备运行环境

    按官方文档确认运行环境(常见为 Node.js 或 Python),装好后再执行安装命令。

  2. 2
    安装服务

    复制上面的安装命令执行;命令里带占位符的位置,换成自己的密钥或路径。

  3. 3
    写入客户端配置

    在客户端的 MCP 配置里填好启动命令和参数,保存后完全退出客户端再打开,让新配置生效。

  4. 4
    验证是否可用

    连接成功后客户端会列出约 3 个工具,随便调一个只读工具确认能返回结果。

典型使用场景

3 类
按分类与标签整理的常见用法,具体以官方文档为准
出图与素材

生成配图、图标与素材。

设计稿处理

读取设计稿、批量改尺寸或导出。

内容生成

按主题生成文案与视觉草稿。

选择建议

选之前先看这几条
按条目字段给出的判断,不替你决定用不用

值得选的地方

  • 资料标注 3 个工具,职责相对集中,接入后不容易误操作。
  • 已适配 Claude Desktop、Claude,换客户端不用重新研究配置。
  • 资料标注无需鉴权,装好就能连。

先想清楚的地方

  • 没有现成安装命令,需要按官方文档自己填启动命令和参数。
  • 版本更新与行为细节以官方仓库为准。

更适合

  • 需要在设计创作场景里让模型直接调用外部能力的人。
  • 已经在用 Claude Desktop、Claude 的用户,接入路径最短。
  • 想先快速试一下、能接受在本地或自有环境跑服务的人。

不太适合

  • 需要厂商级 SLA 和技术支持的生产环境。

同类对比

6 个同类
同分类里挑最接近的几条,字段以各自详情页为准
MCP 服务器工具数传输方式部署方式安装命令最后核验
MCP Image Gen 268bb901当前 3 未标注 未标注 无 2026-05-17
Cursor Talk To Figma MCP Da1c7911 40 未标注 未标注 无 2026-05-17
Blender MCP 17 stdio · 本地进程 本地部署 有 2026-05-18
Nova Canvas MCP Server 2 stdio · 本地进程 本地部署 有 2026-05-17
Figma Context MCP 3aaa8d28 未标注 未标注 支持远程部署 无 2026-05-17
kreuzberg 未标注 stdio · 本地进程 本地部署 有 2026-04-12

服务器档案

设计创作
字段来自项目仓库与公开索引,中文说明由本站标注
提供方aitherlab-dev上游标注的维护方 aitherlab-dev
分类设计创作本站按上游分类展示
连接方式未标注上游未给出传输方式
鉴权方式无需鉴权远程服务建议用最小权限凭据
维护方社区 / 第三方核验等级 仅资料核对
包名 / 运行时未标注 · 未标注本地安装时用到的包标识与运行时
部署类型未标注上游给出的部署形态
最近核验2026-05-17上游最后核验时间

详细介绍

7.9 KB
来自项目仓库与上游目录的原文,未做删改

mcp-image-gen

MCP server for local image generation using stable-diffusion.cpp via diffusion-rs. Runs entirely on your machine — no API keys, no cloud, no data leaves your computer.

Works with any MCP client: Claude Code CLI, Claude Desktop, or any other MCP-compatible tool.

What it does

Gives your AI assistant three tools:

Tool Description
generate_image Generate an image from a text prompt. Returns the file path
download_model Download a model from HuggingFace
list_models List available (downloaded) models

Supported architectures: FLUX.2, FLUX.1, SDXL, and compatible models in GGUF/safetensors format.

Features

  • LoRA support — apply .safetensors LoRA adapters with configurable strength
  • Multiple models — switch between models via models.json config
  • Auto-download — models are downloaded from HuggingFace on first use
  • CUDA acceleration — optional, via feature flag
  • HuggingFace token — supports gated models (FLUX.2-dev, etc.)
  • CPU offloading — run large models on limited VRAM
  • Flash attention & VAE tiling — per-model optimization flags

Quick start

1. Build```bash

git clone https://github.com/aitherlab-dev/mcp-image-gen.git
cd mcp-image-gen
cargo build --release


With CUDA support:```bash
cargo build --release --features cuda

The binary will be at target/release/mcp-image-gen.

Optionally, copy it somewhere on your $PATH:```bash
cp target/release/mcp-image-gen ~/.local/bin/


### 2. Add to Claude Code

Edit `~/.claude.json` (or create it) and add the server to `mcpServers`:```json
{
  "mcpServers": {
    "image-gen": {
      "command": "/home/you/.local/bin/mcp-image-gen",
      "env": {}
    }
  }
}

That's it. The server will use default paths and auto-download the default model on first use.

Custom paths (optional)

If you want to control where models and images are stored:```json
{
"mcpServers": {
"image-gen": {
"command": "/home/you/.local/bin/mcp-image-gen",
"env": {
"MCP_IMAGE_GEN_MODELS_PATH": "/home/you/ai-models",
"MCP_IMAGE_GEN_IMAGES_PATH": "/home/you/ai-images",
"HF_TOKEN": "hf_your_token_here"
}
}
}
}


#### Claude Desktop

Same config format — add to `claude_desktop_config.json`:```json
{
  "mcpServers": {
    "image-gen": {
      "command": "/home/you/.local/bin/mcp-image-gen",
      "env": {}
    }
  }
}

3. Use it

Start Claude Code and ask it to generate an image:```
> Generate an image of a sunset over mountains


On first run, the model will be downloaded automatically (~4.4 GB for the default FLUX.2 Klein 4B). Subsequent runs use the cached model.

The server returns the file path of the generated image. In Claude Code CLI you'll see something like:```
Image saved: /home/you/.local/share/mcp-image-gen/images/20250326_142030_abc12345.png

Configuration

Settings file```

~/.config/mcp-image-gen/settings.json
``````json
{
"modelsPath": "/home/you/.local/share/mcp-image-gen/models",
"imagesPath": "/home/you/.local/share/mcp-image-gen/images",
"selectedModel": "flux2-klein-4b",
"width": 1024,
"height": 1024,
"steps": 20
}


Created automatically with defaults on first run. You can edit it manually or let the server manage it.

### Environment variables

Override any config value via environment:

| Variable | Description | Default |
|----------|-------------|---------|
| `MCP_IMAGE_GEN_MODELS_PATH` | Where to store/cache downloaded models | `~/.local/share/mcp-image-gen/models` |
| `MCP_IMAGE_GEN_IMAGES_PATH` | Where generated images are saved | `~/.local/share/mcp-image-gen/images` |
| `MCP_IMAGE_GEN_SELECTED_MODEL` | Default model ID | `flux2-klein-4b` |
| `HF_TOKEN` | HuggingFace API token for gated models | — |
| `RUST_LOG` | Log level (e.g. `mcp_image_gen=debug`) | `mcp_image_gen=info` |

### Models config

Models are defined in `~/.config/mcp-image-gen/models.json`. Created automatically with a default model on first run.

Each model entry specifies HuggingFace repos/files for all components:```json
[
  {
    "id": "flux2-klein-4b",
    "name": "FLUX.2 Klein 4B",
    "diffusion": { "repo": "leejet/FLUX.2-klein-4B-GGUF", "file": "flux-2-klein-4b-Q8_0.gguf" },
    "vae": { "repo": "black-forest-labs/FLUX.2-dev", "file": "vae/diffusion_pytorch_model.safetensors" },
    "llm": { "repo": "unsloth/Qwen3-4B-GGUF", "file": "Qwen3-4B-Q8_0.gguf" },
    "steps": 4,
    "cfg_scale": 1.0,
    "width": 1024,
    "height": 1024,
    "offload_cpu": true,
    "flash_attn": true,
    "vae_tiling": true
  }
]

Adding custom models

Add entries to models.json with the HuggingFace repo and filename for each component:

Field Required Description
id yes Unique identifier
name yes Display name
diffusion yes Main model file (GGUF or safetensors) — {"repo": "...", "file": "..."}
vae no VAE decoder
llm no Text encoder (FLUX models)
clip_l no CLIP-L text encoder (SDXL)
t5xxl no T5-XXL encoder (FLUX)
single_file no true for single-file models (SDXL safetensors)
steps no Default inference steps (default: 4)
cfg_scale no Classifier-free guidance scale (default: 1.0)
width / height no Default resolution (default: 1024)
offload_cpu no Offload model params to CPU (saves VRAM)
flash_attn no Enable flash attention
vae_tiling no Enable VAE tiling (saves VRAM for large images)

LoRA adapters

Add LoRA fields to any model entry in models.json:```json
{
"id": "flux2-klein-4b",
"lora": "/home/you/loras/anime-style.safetensors",
"lora_strength": 0.8,
"lora_enabled": true
}


| Field | Description |
|-------|-------------|
| `lora` | Full path to a `.safetensors` LoRA file |
| `lora_strength` | Multiplier, 0.0–2.0 (default: 1.0) |
| `lora_enabled` | Toggle on/off without removing the path (default: true) |

## Default model

**FLUX.2 Klein 4B** (Q8_0 GGUF) — ~4.4 GB total:
- Diffusion model: `leejet/FLUX.2-klein-4B-GGUF`
- VAE: `black-forest-labs/FLUX.2-dev`
- Text encoder: `unsloth/Qwen3-4B-GGUF`

Fast (4 steps), good quality at 1024×1024. Runs on 8 GB VRAM with CPU offloading.

## Requirements

- **Rust 1.75+**
- **CMake** (for building stable-diffusion.cpp)
- **CUDA toolkit** (optional, for GPU acceleration)
- ~4.4 GB disk space for the default model
- 8 GB+ VRAM recommended (CPU offloading available for less)

### Linux (Ubuntu/Debian)```bash
sudo apt install cmake build-essential

Linux (Arch)```bash

sudo pacman -S cmake base-devel


### macOS```bash
brew install cmake

How it works```

Claude Code ←→ mcp-image-gen (stdio, JSON-RPC) ←→ diffusion-rs ←→ stable-diffusion.cpp
```

The server communicates via stdin/stdout using the MCP protocol (JSON-RPC over stdio). Image generation runs in a separate thread to not block the protocol handler.

Generated images are saved to disk and the file path is returned as text. Your MCP client (Claude Code, Claude Desktop, etc.) receives the path and can display or reference the image.

Troubleshooting

"Failed to build HF API" — check your internet connection. Models are downloaded from HuggingFace on first use.

"This model requires a HuggingFace token" — set HF_TOKEN in your env or ~/.claude.json. Some models (like FLUX.2-dev) require accepting a license on huggingface.co first.

Out of VRAM — enable offload_cpu: true and vae_tiling: true in your model config. Or use a smaller model.

Slow generation — build with --features cuda for GPU acceleration. Without it, inference runs on CPU.

Logs — stderr output goes to Claude Code's MCP server logs. Set RUST_LOG=mcp_image_gen=debug for verbose output.

License

MIT

Credits

Built by aitherlab. Part of the aitherflow project.

Powered by:
- diffusion-rs — Rust bindings for stable-diffusion.cpp
- stable-diffusion.cpp — C++ inference engine

常见问题

6 问
答案由本站依据条目字段整理,不引用上游原文
MCP Image Gen 268bb901 怎么安装?

资料没有给出安装命令,请到官方仓库或文档按说明接入;这类条目多数是生态应用,或尚未发布可安装包。

MCP Image Gen 268bb901 怎么在 Claude Desktop、Claude 里配置?

它的传输方式未标注,在客户端的 MCP 配置里填好启动命令和参数,保存后完全退出客户端再打开,配置才会生效。

MCP Image Gen 268bb901 收费吗?

本站不代收费用,也没有标注具体价格。是否收费取决于服务本身和背后的接口额度,正式使用前按官方说明确认。

MCP Image Gen 268bb901 需要 API Key 吗?

资料标注为无需认证,装好就能连;如果实际调用时报鉴权错误,按官方文档补上凭据。

MCP Image Gen 268bb901 支持哪些客户端?

资料标注支持 Claude Desktop、Claude;其他客户端只要能配置 MCP Server,也能用同样的命令或地址接入。

MCP Image Gen 268bb901 能提供哪些能力?

资料显示约 3 个工具;能力标签包括 图像生成、本地处理、Rust、Claude、模型管理。具体工具清单以官方文档为准。

热门专题

按标签挑服务器
全部专题 →

服务器信息以作者仓库与官方文档为准;分类、传输方式与鉴权方式的中文说明为中文整理,最近核验 2026-05-17。