ultimate-gemini-mcp

Ultimate Gemini MCP

MCP server for Google's Gemini 3.1 Flash Image — fast image generation with advanced reasoning, 512px–4K resolution, up to 14 reference images, Google Search grounding, and automatic thinking mode.

Setup

Claude Desktop

Add to your mcp config (mcp.json / .claude.json):

{
  "mcpServers": {
    "ultimate-gemini": {
      "command": "uvx",
      "args": ["ultimate-gemini-mcp@latest"],
      "env": {
        "GEMINI_API_KEY": "your-api-key-here"
      }
    }
  }
}

All generated images include invisible SynthID watermarks for authenticity and provenance tracking.

Features

Gemini 3.1 Flash Image

• High-Resolution Output: 512px, 1K, 2K, and 4K resolution
• Advanced Text Rendering: Legible, stylized text in infographics, menus, diagrams, and logos
• Up to 14 Reference Images: Up to 10 objects + 4 characters for style/character consistency
• Google Search Grounding: Real-time data (weather, stocks, events, maps)
• Google Image Search: Visual context from web images — the model can FIND real images of anything
• Thinking Mode: Configurable reasoning - "minimal" (fast) or "high" (best quality)
• Transparent Backgrounds: Flip one flag → ready-to-use transparent PNG/WebP cut-outs with a real alpha channel, recovered by a two-pass difference matte (generate on white → edit to black → solve for alpha). True soft edges/glow/glass, no color halo. Pillow only — no extra dependencies. Costs a second model call (~2x).
• Dedicated App-Icon / Logo Tool: generate_app_icon forces a square, transparent, 1024px PNG every time — no way to get a non-square or opaque-background icon

This model is different. Unlike traditional image generators that rely solely on training data, Gemini 3.1 Flash has live access to Google Search and Image Search. It can find actual references for products, people, events, or anything that exists online. "Way of Wade 12" → generates the REAL shoe. "Tony Hawk" → finds real photos. Don't over-prompt — let the model cook.

Server Features

• Batch Processing: Generate multiple images in parallel (up to 8 concurrent)
• 26 Expert Prompt Templates: MCP slash commands for photography, cinematics, storyboards, and more
• Flexible Aspect Ratios: 14 options — 1:1, 1:4, 1:8, 2:3, 3:2, 3:4, 4:1, 4:3, 4:5, 5:4, 8:1, 9:16, 16:9, 21:9
• Configurable via Environment Variables: Output directory, default size, timeouts, and more

Showcase

Photorealistic Capabilities

Jensen Huang — GPU Surfing

Elon Musk — Mars Chess Match

Jensen Huang — GPU Kitchen

Elon Musk — Cybertruck Symphony

Jensen Huang — Underwater Data Center

Elon Musk — SpaceX Skateboarding

Google Search Grounding

Current Weather in San Francisco

Google Image Search

Butterfly on Flower

Different Resolutions

512px (fastest)

1K

2K

Quick Start

Prerequisites

• Python 3.11+
• Google Gemini API key (free tier available)

Installation

Using uvx (recommended — no install needed):

uvx ultimate-gemini-mcp@latest

Note: Use @latest to ensure uv always fetches the newest version from PyPI. Without it, uv may use a cached environment.

Using pip:

pip install ultimate-gemini-mcp

From source:

git clone https://github.com/anand-92/ultimate-image-gen-mcp
cd ultimate-image-gen-mcp
uv sync

Config file locations:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json

macOS spawn uvx ENOENT error: Use the full path — find it with which uvx, then set "command": "/Users/you/.local/bin/uvx".

Claude Code

claude mcp add ultimate-gemini \
  --env GEMINI_API_KEY=your-api-key \
  -- uvx ultimate-gemini-mcp@latest

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "ultimate-gemini": {
      "command": "uvx",
      "args": ["ultimate-gemini-mcp@latest"],
      "env": {
        "GEMINI_API_KEY": "your-api-key-here"
      }
    }
  }
}

Images are saved to ~/gemini_images by default. Add "OUTPUT_DIR": "/your/path" to customize.

Tools

generate_image

Generate an image with Gemini 3.1 Flash Image.

Parameter	Type	Default	Description
prompt	string	required	Text description. Less is more — "Tony Hawk kickflip" beats a long description. The model with search can find references automatically.
aspect_ratio	string	1:1	One of: 1:1 1:4 1:8 2:3 3:2 3:4 4:1 4:3 4:5 5:4 8:1 9:16 16:9 21:9
image_size	string	2K	512px, 1K, 2K, or 4K
output_format	string	png	png, jpeg, or webp
reference_image_paths	list	[]	Up to 14 local image paths (10 objects + 4 characters)
enable_google_search	bool	false	USE THIS for products, people, events — anything that exists now. The model searches Google for real info.
enable_image_search	bool	false	USE THIS for visual references. The model finds actual images to work from. This is huge — it can reference real photos of anyone/anything.
thinking_level	string	minimal	minimal (fast) or high (best quality)
response_modalities	list	["TEXT","IMAGE"]	["TEXT","IMAGE"], ["IMAGE"], or ["TEXT"]
transparent_background	bool	false	Produce a transparent PNG/WebP cut-out via the two-pass difference matte (~2x cost; see below)
preserve_original	bool	true	Also keep the pass-1 (white-background) image, not just the cut-out
alpha_output_format	string	png	Alpha-capable output format: png or webp

Image size guide:

• 512px — fastest, lowest cost (0.5K)
• 1K — fast, good for testing (~1-2 MB)
• 2K — recommended for most use cases (~3-5 MB)
• 4K — maximum quality for production assets (~8-15 MB)

Transparent backgrounds — set one flag, get a real alpha cut-out

Just set transparent_background=true. You get back a ready-to-use transparent PNG/WebP (real alpha channel) at transparent_path — no manual masking, no second tool, no follow-up steps.

Generating an app icon or logo? Use the dedicated generate_app_icon tool instead — it forces square + transparent + 1024px PNG so the icon constraints can't be set wrong.

Under the hood this is a two-pass difference matte. The subject is rendered once on a pure white (#FFFFFF) background, that image is edited to a pure black (#000000) background, and the two frames are combined to solve for alpha per pixel: since obs_white − obs_black = (1−α)·255 on every channel, α = 1 − mean(obs_white − obs_black)/255, and the foreground colour is un-premultiplied from the black frame. Because there's no colour key, there's no green spill/halo; alpha is fractional, so soft edges, glow, glass, and faint shadows all survive. Pillow-only, zero ML downloads — but it costs a second model call (~2x tokens/latency).

The technique assumes the edit pass changed only the background. If the model drifts the subject between passes, the matte degrades — the result still returns (aligned/alignment_error flag it, with a loud post_processing_warnings entry) so you can decide whether to regenerate.

Each returned image gains: transparent_path, background_removed, background_removal_mode ("difference_matte"), aligned, alignment_error, alpha_output_format, and post_processing_warnings. By default the pass-1 (white-background) original is preserved alongside the cut-out (preserve_original=true).

// generate_image(prompt="a friendly robot mascot", transparent_background=true)
{
  "images": [{
    "path": "/path/to/a-friendly-robot-mascot-...png",            // pass-1 (white bg)
    "transparent_path": "/path/to/a-friendly-robot-mascot-...-transparent.png",
    "background_removed": true,
    "background_removal_mode": "difference_matte",
    "aligned": true,
    "alignment_error": 0.004,
    "alpha_output_format": "png",
    "post_processing_warnings": []
  }]
}

It nails crisp-edged subjects and soft glow/glass. The one failure mode is the edit pass drifting the subject (flagged via aligned: false) — regenerate if edges look ghosted.

generate_app_icon

Purpose-built for app icons and logos. Square, transparent, and 1024px are forced — there is no aspect_ratio, image_size, output_format, or transparent_background knob to get wrong. Every result is a real alpha-channel PNG at transparent_path, ready to drop into a .iconset directory and convert with iconutil -c icns.

Parameter	Type	Default	Description
prompt	string	required	Describe the icon/logo mark only — framing & transparency are handled
reference_image_paths	str | list	null	Brand/style reference image path(s), up to 14
enable_google_search	bool	false	Ground design in real web references
enable_image_search	bool	false	Use Google Image Search for visual context
thinking_level	string	high	minimal or high (icons reward high)
allow_icon_words_in_prompt	bool	false	Escape hatch — bypass the prompt guard only when a word like "logo" is genuinely part of the subject

The prompt must describe ONLY the subject, never the deliverable. This tool already turns whatever you describe into an icon, so framing words like "app icon", "logo", "favicon", or "squircle" in the prompt are rejected (set allow_icon_words_in_prompt=true only if such a word is literally part of the depicted subject). Right: "a glowing electric-blue magnifying glass over a network graph". Wrong: "an app icon of a magnifying glass".

// generate_app_icon(prompt="a glowing electric-blue magnifying glass over a network graph")
{
  "images": [{
    "transparent_path": "/path/to/...-transparent.png",  // square, 1024px, alpha
    "background_removed": true,
    "alpha_output_format": "png"
  }]
}

batch_generate

Generate multiple images in parallel.

Parameter	Type	Default	Description
prompts	list	required	List of prompt strings (max 8)
aspect_ratio	string	1:1	Aspect ratio applied to all images
image_size	string	2K	Resolution for all images
output_format	string	png	Format for all images
response_modalities	list	["TEXT","IMAGE"]	Modalities for all images
batch_size	int	8	Max concurrent requests
enable_image_search	bool	false	Use Google Image Search for visual context
thinking_level	string	minimal	minimal or high
transparent_background	bool	false	Apply the two-pass difference matte to every image (each costs a second model call)
preserve_original	bool	true	Keep the pass-1 (white-background) images too
alpha_output_format	string	png	Transparent output format: png or webp

MCP Prompt Templates

26 expert prompt templates are available as MCP slash commands in Claude Code (type / to browse). Each template returns a crafted prompt and recommended parameters ready to pass directly to generate_image or batch_generate. For app icons and logos, use the dedicated generate_app_icon tool instead.

Command	Description	Default aspect ratio
photography_shot	Photorealistic shot with lens/lighting specs	16:9
cinematic_scene	Film still with cinematography language	21:9
product_mockup	Commercial e-commerce photography	1:1 or 4:5
batch_storyboard	Multi-scene storyboard → calls batch_generate	16:9
macro_shot	Extreme macro with micro-snoot lighting	1:1
fashion_portrait	Editorial fashion with gobo shadow patterns	4:5
technical_cutaway	Stephen Biesty-style cutaway diagram	3:2, 4K, IMAGE only
flat_lay	Overhead knolling photography	1:1
action_freeze	High-speed strobe with motion blur background	16:9
night_street	Moody night street with practical light sources	16:9
drone_aerial	Straight-down golden hour aerial	4:5, 4K, IMAGE only
stylized_3d_render	UE5-style render with subsurface scattering	1:1, IMAGE only
sem_microscopy	Scanning electron microscope false-color	1:1, IMAGE only
double_exposure	Silhouette-blended double exposure	2:3, IMAGE only
architectural_viz	Ray-traced architectural visualization	3:2, 4K
isometric_illustration	Orthographic isometric 3D illustration	1:1, IMAGE only
food_photography	High-end backlit food photography	4:5
motion_blur	Rear-curtain sync slow shutter sequence	16:9
typography_physical	Text embedded in physical environment	16:9, 4K, IMAGE only
retro_futurism	1970s cassette-futurism analog sci-fi	4:3, IMAGE only
surreal_dreamscape	Surrealist impossible physics scene	1:1, IMAGE only
character_sheet	Video game character concept art sheet	3:2, 4K, IMAGE only
pbr_texture	Seamless PBR texture map with raking light	1:1, IMAGE only
historical_photo	Period-accurate photography with film emulation	4:5
bioluminescent_nature	Long-exposure bioluminescence macro	1:1
silhouette_shot	Cinematic pure-black silhouette master shot	21:9, 4K

Configuration

Variable	Default	Description
GEMINI_API_KEY	—	Required. Google Gemini API key
OUTPUT_DIR	~/gemini_images	Directory where images are saved
DEFAULT_IMAGE_SIZE	2K	Default resolution (1K, 2K, 4K)
DEFAULT_MODEL	gemini-3-pro-image-preview	Default model
ENABLE_PROMPT_ENHANCEMENT	false	Auto-enhance prompts by default
ENABLE_GOOGLE_SEARCH	false	Enable Google Search grounding by default
REQUEST_TIMEOUT	60	API timeout in seconds
MAX_BATCH_SIZE	8	Max parallel requests in batch mode
LOG_LEVEL	INFO	Logging level

Troubleshooting

spawn uvx ENOENT — Claude Desktop can't find uvx. Use the full path:

"command": "/Users/yourusername/.local/bin/uvx"

Find it with: which uvx

GEMINI_API_KEY not found — Set the key in your MCP config env block or in a .env file. Get a free key at Google AI Studio.

Content blocked by safety filters — Rephrase the prompt to avoid sensitive content.

Rate limit exceeded — Wait and retry, or upgrade your API quota.

Images not saving — Check OUTPUT_DIR exists and is writable: mkdir -p /your/output/path.

License

MIT — see LICENSE for details.

MCP Registry

mcp-name: io.github.nikships/ultimate-image-gen-mcp

Links

• Google AI Studio — Get your API key
• Gemini API Docs
• Model Context Protocol
• FastMCP