creative-toolkit（创意工具包）

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@1.2.13"]
    }
  }
}

Free tools (search, enhance, inspire) work immediately — no API key needed:

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

To unlock image generation, configure one of these providers:

Set credentials in ~/.clawdbot/.env, ~/.config/meigen/config.json, or add an "env" block to the mcporter config above. See references/providers.md for details.

Sandboxed environments: if the host can't write to the default ~/Pictures/meigen save path, set the MEIGEN_OUTPUT_DIR env var to a writable directory (supports ~ expansion).

Available Tools

Free — no API key required

Requires configured provider

Important Rules

Never describe generated images

You cannot see generated images. After generation, only present the exact data from the tool response:

**Direction 1: Modern Minimal**
- Image URL: https://images.meigen.art/...
- Saved to: ~/Pictures/meigen/2026-02-08_xxxx.jpg

Do NOT write creative commentary about what the image "looks like".

Never specify model or provider

Do NOT pass model or provider to generate_image unless the user explicitly asks. The server auto-selects the best available provider and model.

Midjourney V7 vs Niji 7

Both take ~60s, accept 1 reference image, and return 4 candidate images per generation. Advanced params (stylize/chaos/weird/raw/iw/sw/sv) run with fixed server-side defaults and cannot be tuned from MCP — the only exception is sref, settable via --sref <code> at the end of the prompt (Midjourney style codes only, e.g. 3799554500; no URLs or local paths). They differ in content focus and prompt enhancement style:

• Midjourney V7 (model: "midjourney-v7") — general / photorealistic. Use for product photography, portraits, landscapes, cinematic and editorial shots. When enhancing, use style: 'realistic' (the default).
• Midjourney Niji 7 (model: "midjourney-niji7") — anime / illustration ONLY. Do NOT use for photorealistic, product, or non-anime content — use GPT Image 2.0 or Nanobanana 2 instead. The server auto-appends anime illustration style if your prompt lacks anime keywords. When enhancing, ALWAYS pass style: 'anime' to enhance_prompt — the default realistic produces prompts poorly suited for anime models.

Always confirm before generating multiple images

When the user wants multiple variations, present options first and ask which direction(s) to try. Include an "all of the above" option. Never auto-generate all variants without user confirmation.

Workflow Modes

Mode 1: Single Image

User wants one image. Write a prompt (or call enhance_prompt if the description is brief), generate, present URL + path.

Mode 2: Prompt Enhancement + Generation

For brief ideas (under ~30 words, lacking visual details), enhance first:

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

2. generate_image prompt="<enhanced prompt>"
   -> Omit aspectRatio to let MeiGen auto-infer (recommended). Pass an explicit
      value like aspectRatio="16:9" only when the user asked for that ratio.

Mode 3: Parallel Generation (2+ images)

User needs multiple variations — different directions, styles, or concepts.

1. Plan directions, present as a table
2. Ask user which direction(s) to try
3. Write distinct prompts for each — don't just tweak one word
4. Generate selected directions (max 4 parallel for API providers, 1 at a time for ComfyUI)
5. Present URLs + paths

Mode 4: Multi-Step Creative (base + extensions)

User wants a base design plus derivatives (e.g., "design a logo and make mockups").

1. Plan 3-5 directions, ask user which to try
2. Generate selected direction(s)
3. Present results, ask user to approve or try another
4. Plan extensions using the approved Image URL as referenceImages
5. Generate extensions

Never jump from plan to generating everything at once.

Mode 5: Edit/Modify Existing Image

User provides an image and asks for changes (add text, change background, etc.).

• Pass the image (URL or local path) as referenceImages, then generate with a short, literal prompt describing ONLY the edit
• The reference image carries all visual context — do NOT re-describe the original image
• Example prompt: "Add the text 'meigen.ai' at the bottom of this image"

Mode 6: Inspiration Search

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

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

Mode 7: Reference Image Generation

Use an existing image to guide visual style. Pass URLs or local file paths directly to referenceImages.

generate_image prompt="coffee mug mockup with this logo" referenceImages=["~/Desktop/my-logo.png"]
   -> Local files are auto-compressed (max 2MB, 2048px) and prepared for the selected provider

Reference image sources: gallery URLs, previous generation URLs, or local file paths. All providers accept local paths — they are automatically handled.

Mode 8: ComfyUI Workflows

1. comfyui_workflow action="list"           -> See saved workflows
2. comfyui_workflow action="view" name="txt2img"  -> See adjustable parameters
3. comfyui_workflow action="modify" name="txt2img" modifications={"steps": 30}
4. generate_image prompt="..." workflow="txt2img"  -> Generate

Alternative Providers

You can use your own OpenAI-compatible API or a local ComfyUI instance instead of — or alongside — the default MeiGen provider. See references/providers.md for detailed configuration and provider comparison. For MeiGen model pricing, see https://www.meigen.ai/model-comparison.

Troubleshooting

See references/troubleshooting.md for common issues, solutions, and security & privacy details.