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.

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": []
  }]
}