gemini-image-studio-mcp

MCP server for AI image generation and editing with Google Gemini. Create web assets, ad creatives, and brand visuals — with structured JSON editing for precise, repeatable control.

What Makes This Different

Most Gemini image MCP servers are basic text-to-image wrappers. This one adds a structured editing pipeline:

1. Generate an image from text or JSON prompts
2. Decompose it into a structured JSON blueprint (every visual component mapped)
3. Edit by changing specific fields — subject[0].hair.color: "platinum_blonde" — and regenerating

This means precise, isolated changes without affecting the rest of the image. Change a hair color without touching the background. Swap clothing without altering the pose. All through dot-notation JSON paths.

Features

• 5 MCP Tools — generate, decompose, edit, presets, list
• Structured JSON Editing — decompose images into blueprints, edit specific fields with dot-notation
• Natural Language Editing — or just describe the change in plain English
• 10 Built-in Presets — Facebook ads, Instagram stories, hero images, OG images, YouTube thumbnails, and more
• Reference Image Support — up to 14 reference images for character/object consistency
• Dual Model Support — Gemini 3.1 Flash (fast) or Gemini 3 Pro (best quality)
• Blueprint Caching — decomposed blueprints cached alongside images for instant re-edits
• Google Search Grounding — real-world accuracy via web search
• Smart Error Handling — retry on rate limits, clear safety block messages, file size warnings

Quick Start

1. Get a Gemini API Key

Get one free at Google AI Studio.

2. Install

npm install -g gemini-image-studio-mcp

3. Add to Claude Code

claude mcp add gemini-image-studio-mcp -e GEMINI_API_KEY=your-key-here -- gemini-image-studio-mcp

Or add to your project's .claude/mcp.json:

{
  "mcpServers": {
    "gemini-image-studio-mcp": {
      "command": "npx",
      "args": ["-y", "gemini-image-studio-mcp"],
      "env": {
        "GEMINI_API_KEY": "your-key-here"
      }
    }
  }
}

4. Use It

Ask Claude to generate images:

"Create a Facebook ad for a coffee shop with warm lighting"

"Generate a hero image for a tech startup landing page"

"Edit the hero image — change the background to a sunset beach"

Tools

generate_image

Create a new image from text or structured JSON prompts.

Parameter	Type	Required	Description
prompt	string	Yes	Text description or JSON prompt
prompt_format	"text" | "json"	No	Prompt format (default: "text")
preset	string	No	Asset preset (e.g., "facebook_ad", "hero_image")
aspect_ratio	string	No	Override ratio ("1:1", "16:9", "9:16", etc.)
image_size	"1K" | "2K" | "4K"	No	Resolution (default: "1K")
model	"flash" | "pro"	No	Gemini model (default: "flash")
reference_images	string[]	No	Paths to reference images for consistency
output_name	string	No	Custom filename
enable_search_grounding	boolean	No	Use Google Search for accuracy

decompose_image

Analyze an image into a structured JSON blueprint — the first step of the edit workflow.

Parameter	Type	Required	Description
image_path	string	Yes	Path to the image
detail_level	"basic" | "detailed" | "exhaustive"	No	Granularity (default: "detailed")

Returns a full blueprint with subject, scene, technical, composition, text_rendering, style_modifiers, and meta sections — each field precisely describing the image's visual components.

edit_image

Edit an image using JSON changes or natural language.

Parameter	Type	Required	Description
image_path	string	Yes	Path to the image
edit_type	"json" | "natural_language"	Yes	Edit mode
changes	object	For JSON edits	Dot-notation paths to change
instruction	string	For NL edits	Natural language instruction
blueprint	object	No	Blueprint (auto-loaded from cache if omitted)
model	"flash" | "pro"	No	Model (default: "flash")
output_name	string	No	Custom filename

JSON edit example — change hair color and add sunglasses:

{
  "image_path": "/output/portrait.png",
  "edit_type": "json",
  "changes": {
    "subject[0].hair.color": "platinum_blonde",
    "subject[0].accessories": [
      { "item": "sunglasses", "material": "metal", "color": "#C0C0C0" }
    ]
  }
}

Natural language edit example:

{
  "image_path": "/output/portrait.png",
  "edit_type": "natural_language",
  "instruction": "Change the background to a tropical beach at sunset. Keep the person exactly the same."
}

get_presets

List available asset presets with dimensions, tips, and conventions.

Parameter	Type	Required	Description
category	"ad" | "web" | "social" | "all"	No	Filter (default: "all")

list_generated

Browse previously generated images.

Parameter	Type	Required	Description
filter	string	No	Search by filename
limit	number	No	Max results (default: 20)
include_blueprints	boolean	No	Include cached blueprints

JSON Editing Workflow

The key differentiator — precise, field-level image editing:

Step 1: Generate
  generate_image(prompt: "Professional headshot, navy blazer", preset: "linkedin_post")
  → /output/headshot.png

Step 2: Decompose
  decompose_image(image_path: "/output/headshot.png")
  → JSON blueprint with every visual component mapped

Step 3: Edit (precise)
  edit_image(
    image_path: "/output/headshot.png",
    edit_type: "json",
    changes: {
      "subject[0].clothing[0].color": "#8B0000",
      "scene.lighting.type": "studio_softbox"
    }
  )
  → /output/headshot-edit-1.png (blazer changed to dark red, lighting adjusted)

Step 4: Edit (creative)
  edit_image(
    image_path: "/output/headshot-edit-1.png",
    edit_type: "natural_language",
    instruction: "Add warm bokeh to the background"
  )
  → /output/headshot-edit-1-edit-1.png

Dot-Notation Paths

subject[0].hair.color          → Hair color
subject[0].hair.style          → Hair style
subject[0].clothing[0].color   → First clothing item color
subject[0].accessories         → Add/change accessories
scene.lighting.type            → Lighting type
scene.location                 → Location/background
text_rendering.text_content    → Text in image
technical.lens                 → Camera lens
composition.framing            → Shot framing
style_modifiers.aesthetic      → Aesthetic style

Built-in Presets

Preset	Category	Aspect Ratio	Dimensions	Best For
facebook_ad	Ad	1:1	1080x1080	Facebook/Instagram feed ads
instagram_story_ad	Ad	9:16	1080x1920	Instagram/Facebook story ads
google_display_banner	Ad	16:9	1200x628	Google Display Network
hero_image	Web	21:9	2560x1080	Above-the-fold hero sections
og_image	Web	16:9	1200x630	Social share / link previews
product_card	Web	4:5	800x1000	E-commerce product grids
email_header	Web	3:1	600x200	Email marketing headers
linkedin_post	Social	1:1	1080x1080	LinkedIn feed posts
twitter_post	Social	16:9	1200x675	Twitter/X posts
youtube_thumbnail	Social	16:9	1280x720	YouTube thumbnails

Configuration

Variable	Required	Default	Description
GEMINI_API_KEY	Yes	—	Google AI Studio API key
OUTPUT_DIR	No	./output	Where generated images are saved

Integration

Claude Code

claude mcp add gemini-image-studio-mcp -e GEMINI_API_KEY=your-key -- gemini-image-studio-mcp

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "gemini-image-studio-mcp": {
      "command": "npx",
      "args": ["-y", "gemini-image-studio-mcp"],
      "env": {
        "GEMINI_API_KEY": "your-key-here"
      }
    }
  }
}

Any MCP Client

GEMINI_API_KEY=your-key npx gemini-image-studio-mcp

The server communicates over stdio using the Model Context Protocol.

MCP Prompt & Resource

This server also exposes:

• Prompt: nano_banana_expert — invoke this to give Claude full knowledge of the JSON schema, editing best practices, and asset creation guidelines
• Resource: nanobanana://schema/prompt — the raw JSON schema with all enum values for programmatic access

Models

Model	ID	Best For
Flash (default)	gemini-3.1-flash-image-preview	Fast generation, high volume, cost-effective
Pro	gemini-3-pro-image-preview	Best quality, complex scenes, professional assets

Contributing

Contributions welcome! Please:

1. Fork the repository
2. Create a feature branch (git checkout -b feature/my-feature)
3. Run tests (npm test)
4. Commit your changes
5. Push and open a PR

License

MIT