Creative Toolkit

Generate professional AI images through a unified interface that routes across multiple providers. Search curated prompts, enhance ideas into production-ready descriptions, and manage local ComfyUI workflows — all from a single MCP server.

Quick Start

Add the MCP server to your mcporter config (~/.config/mcporter/config.json):

{
  "mcpServers": {
    "creative-toolkit": {
      "command": "npx",
      "args": ["-y", "meigen@latest"],
      "env": {
        "MEIGEN_API_TOKEN": "${MEIGEN_API_TOKEN}"
      }
    }
  }
}

Set your API token in ~/.clawdbot/.env or shell environment:

export MEIGEN_API_TOKEN="meigen_sk_..."

Generate your first image:

mcporter call creative-toolkit.generate_image prompt="a minimalist perfume bottle on white marble, soft directional lighting, product photography"

Or try it without any config (ad-hoc stdio mode):

mcporter call --stdio "npx -y meigen@latest" generate_image prompt="a ceramic vase with morning light"

No API key? Free tools still work:

mcporter call creative-toolkit.search_gallery query="cyberpunk"
mcporter call creative-toolkit.enhance_prompt brief="a cat in space" style="realistic"

Setup

Get an API Token

1. Visit meigen.ai → sign in → click avatar → Settings → API Keys
2. Create a new key (starts with meigen_sk_)
3. Set as environment variable or save to config:

# Shell environment or ~/.clawdbot/.env
export MEIGEN_API_TOKEN="meigen_sk_..."

Or save to ~/.config/meigen/config.json:

{
  "meigenApiToken": "meigen_sk_..."
}

Alternative Providers

You can use your own OpenAI-compatible API or a local ComfyUI instance instead of — or alongside — the default provider. Save to ~/.config/meigen/config.json:

OpenAI / Together AI / Fireworks AI:

{
  "openaiApiKey": "sk-...",
  "openaiBaseUrl": "https://api.together.xyz/v1",
  "openaiModel": "black-forest-labs/FLUX.1-schnell"
}

Local ComfyUI:

{
  "comfyuiUrl": "http://localhost:8188"
}

Import workflows with the comfyui_workflow tool (action: import). The server auto-detects key nodes (KSampler, CLIPTextEncode, EmptyLatentImage) and fills in prompt, seed, and dimensions at runtime.

Multiple providers can be configured simultaneously. Auto-detection priority: MeiGen > ComfyUI > OpenAI.

Available Tools

Free — no API key required

Tool	What it does
search_gallery	Semantic search across AI image prompts — finds conceptually similar results, not just keyword matches. Also supports category filtering and curated browsing. Returns prompt text, thumbnails, and metadata.
get_inspiration	Get the full prompt and high-res images for any gallery entry. Use after search_gallery to get copyable prompts.
enhance_prompt	Expand a brief idea (e.g. "a cat in space") into a detailed, style-aware prompt with lighting, composition, and material directions. Supports three styles: realistic, anime, illustration.
list_models	List all available models across configured providers with capabilities and supported features.

Requires configured provider

Tool	What it does
generate_image	Generate an image from a text prompt. Automatically routes to the best available provider. Supports aspect ratio, seed, and reference images.
upload_reference_image	Compress and upload a local image (max 2MB, 2048px) for use as a style reference in generation. Returns a public URL.
comfyui_workflow	List, view, import, modify, and delete ComfyUI workflow templates. Modify parameters like steps, CFG scale, sampler, and checkpoint without editing JSON.

Usage Patterns

Basic generation

Generate a product photo of a ceramic vase with morning light

The server picks the best provider, generates the image, and returns a URL + local file path.

Prompt enhancement then generation

For brief ideas, enhance first for much better results:

1. enhance_prompt brief="futuristic city" style="realistic"
   → Returns detailed prompt with camera lens, lighting setup, atmospheric effects

2. generate_image prompt="<enhanced prompt>" aspectRatio="16:9"
   → Generates with the enhanced prompt

Reference image workflow

Use an existing image to guide style and composition:

1. upload_reference_image filePath="~/Desktop/my-logo.png"
   → Returns public URL

2. generate_image prompt="coffee mug mockup with this logo" referenceImages=["<url>"]
   → Generates using the reference for style guidance

Reference images work across all providers.

Gallery exploration

Semantic search understands intent — "dreamy portrait with soft light" finds relevant results even without exact keyword matches:

1. search_gallery query="dreamy portrait with soft light"
   → Finds semantically similar prompts with thumbnails

2. search_gallery category="Product"
   → Browse by category from 1,300+ curated prompts

3. get_inspiration id="<entry_id>"
   → Get full prompt text — copy and modify for your own generation

ComfyUI workflows

1. comfyui_workflow action="list"
   → See saved workflows

2. comfyui_workflow action="view" name="txt2img"
   → See adjustable parameters (steps, CFG, sampler, checkpoint)

3. comfyui_workflow action="modify" name="txt2img" modifications={"steps": 30, "cfg": 7.5}
   → Adjust without editing JSON

4. generate_image prompt="..." workflow="txt2img"
   → Generate using the custom workflow

Provider Comparison

	MeiGen Platform	OpenAI-Compatible	ComfyUI (Local)
Models	Nanobanana 2, Seedream 5.0, GPT Image 1.5, etc.	Any model at the endpoint	Any checkpoint on your machine
Reference images	Native support	gpt-image-1.5 only	Requires LoadImage node
Concurrency	Up to 4 parallel	Up to 4 parallel	1 at a time (GPU constraint)
Latency	10-30s typical	Varies by provider	Depends on hardware
Cost	Token-based credits	Provider billing	Free (your hardware)
Offline	No	No	Yes

MeiGen Model Pricing

Model	Credits	4K	Best For
Nanobanana 2 (default)	5	Yes	General purpose, high quality
Seedream 5.0 Lite	5	Yes	Fast, stylized imagery
GPT Image 1.5	2	No	Budget-friendly
Nanobanana Pro	10	Yes	Premium quality
Seedream 4.5	5	Yes	Stylized, wide ratio support
Midjourney Niji 7	15	No	Anime and illustration

When no model is specified, the server defaults to Nanobanana 2.

Prompt Enhancement Styles

enhance_prompt supports three style modes, each producing different types of detail:

Style	Focus	Best For
realistic	Camera lens, aperture, focal length, lighting direction, material textures	Product photos, portraits, architecture
anime	Key visual composition, character details (eyes, hair, costume), trigger words	Anime illustrations, character design
illustration	Art medium, color palette, composition direction, brush texture	Concept art, digital painting, watercolor

Security & Privacy

Remote package execution: This skill runs as an MCP server via npx meigen@latest. The package is published on npmjs.com under the meigen name with full source code available at GitHub. No code is obfuscated or minified beyond standard TypeScript compilation.

Reference image upload: The upload_reference_image tool compresses local images (max 2MB, 2048px) and uploads them to a CDN for use as generation references. This is always user-initiated — the tool never accesses or uploads files without explicit invocation. Uploaded images are used solely as style references for image generation.

API tokens: MEIGEN_API_TOKEN is stored locally in environment variables or ~/.config/meigen/config.json with chmod 600 permissions. Tokens are only sent to the configured provider's API endpoint and never logged or transmitted elsewhere.

No telemetry: The MCP server does not collect analytics, usage data, or send any information to third parties beyond the configured image generation provider.

Troubleshooting

"No image generation providers configured" → Set MEIGEN_API_TOKEN or configure an alternative provider in ~/.config/meigen/config.json

Timeout during generation → Image generation typically takes 10-30 seconds. During high demand, it may take longer. The server polls with a 5-minute timeout.

ComfyUI connection refused → Ensure ComfyUI is running and accessible at the configured URL. Test with: curl <url>/system_stats

"Model not found" → Run list_models to see available models for your configured providers.

Reference image rejected → Images must be public URLs (not local paths). Use upload_reference_image to convert local files to URLs first.