Agileimagegen

Thin MCP server for Gemini image generation and image editing using a Google AI Studio API key.

This project is designed for fast AI-assisted iteration. For repo-specific guidance, see AGENTS.md, PROJECT_STATE.md, and ARCHITECTURE.md.

What It Does

• Exposes exactly 2 MCP tools: image.generate and image.edit
• Uses @google/genai with GOOGLE_API_KEY
• Runs as a local stdio MCP server
• Can also run from Docker with the same env contract
• Saves generated images to disk and returns structured metadata
• Supports reference-guided generation with anchor images
• Uses a shared transparency pipeline across generate and edit

Requirements

• Node.js 20+
• A Google AI Studio API key with access to Gemini image-capable models

Environment

Copy .env.example to .env and fill in your key:

GOOGLE_API_KEY=your-google-ai-studio-api-key
AGILEIMAGEGEN_DEFAULT_MODEL=gemini-2.5-flash-image
AGILEIMAGEGEN_OUTPUT_DIR=./output
AGILEIMAGEGEN_LOG_LEVEL=info
AGILEIMAGEGEN_SAVE_PROMPTS=false

Notes:

• GOOGLE_API_KEY is required.
• AGILEIMAGEGEN_DEFAULT_MODEL can be overridden per tool call.
• AGILEIMAGEGEN_OUTPUT_DIR is where generated images are written by default.
• .env is gitignored and should stay local.

Local Development

Install dependencies:

cmd /c npm install

Run in dev mode:

cmd /c npm run dev

Build:

cmd /c npm run build

Run the built server:

cmd /c npm start

Run tests:

cmd /c npm test

Run live smoke tests:

cmd /c npm run smoke:generate
cmd /c npm run smoke:edit

Docker

Build:

docker build -t agileimagegen-mcp .

Run:

docker run --rm -i --env-file .env -v "${PWD}/output:/app/output" agileimagegen-mcp

The container expects to run as a stdio MCP server, so use -i and wire it through your MCP client.

MCP Client Example

Example local stdio MCP config:

{
  "mcpServers": {
    "agileimagegen": {
      "command": "node",
      "args": ["C:/git/agileimagegen-mcp/dist/server.js"],
      "cwd": "C:/git/agileimagegen-mcp",
      "env": {
        "GOOGLE_API_KEY": "your-key-here"
      }
    }
  }
}

If you prefer .env, keep the cwd pointed at this repo so the server can load it locally.

Tools

image.generate

Input:

{
  "prompt": "Arcade grime sewer cartoon logo",
  "model": "gemini-2.5-flash-image",
  "reference_image_paths": ["C:/temp/input/anchor-logo.png"],
  "size": "square",
  "background": "transparent",
  "transparency_mode": "repair",
  "transparency_threshold": "balanced",
  "filename_hint": "sewer-logo",
  "output_dir": "C:/temp/output"
}

Supported size inputs:

• preset: square, landscape, portrait, widescreen
• explicit: WIDTHxHEIGHT
• or width + height

Transparency controls:

• transparency_mode: off, validate, or repair
• transparency_threshold: balanced or strict

Reference guidance:

• reference_image_paths: optional local anchor images used to steer image.generate
• when present, generate requests are sent as multimodal requests instead of text-only prompts

Defaults:

• background: "transparent" implies transparency_mode: "repair"
• otherwise transparency handling defaults to off
• transparent workflows prefer a chroma-key background color of #01FF01, but can also accept good native alpha or infer and remove a different solid border color when the provider drifts
• image.edit and image.generate both run through the same transparency validation/extraction pipeline

image.edit

Input:

{
  "prompt": "Make this sign grimier and add a toxic green edge glow",
  "input_image_paths": ["C:/temp/input/sign.png"],
  "model": "gemini-2.5-flash-image",
  "transparency_mode": "repair",
  "transparency_threshold": "balanced",
  "filename_hint": "sign-edit",
  "output_dir": "C:/temp/output"
}

For image.edit, transparency repair runs by default when the prompt implies transparent or alpha output.

Tool Output Shape

Both tools return structured content in this shape:

{
  "path": "C:/git/agileimagegen-mcp/output/123456-sewer-logo.png",
  "mime_type": "image/png",
  "model": "gemini-2.5-flash-image",
  "provider": "google",
  "prompt_summary": "Arcade grime sewer cartoon logo",
  "warnings": [],
  "width": 1024,
  "height": 1024,
  "transparency": {
    "requested": true,
    "mode": "repair",
    "threshold": "balanced",
    "source_mime_type": "image/jpeg",
    "has_alpha": true,
    "alpha_pixel_ratio": 0.44,
    "fully_transparent_ratio": 0.39,
    "opaque_border_ratio": 0.02,
    "checkerboard_detected": false,
    "key_color": "#01FF01",
    "key_color_match_ratio": 0.91,
    "background_mode": "keyed",
    "repair_attempted": true,
    "repair_succeeded": true,
    "warnings": []
  }
}

Design Notes

• Width, height, size, and transparent background are passed as prompt guidance because Gemini image-capable models may not honor them as hard output controls in all cases.
• When transparency is requested, the server uses a tiered strategy: accept usable native alpha first, otherwise prefer the requested #01FF01 chroma-key background, then fall back to inferring and removing a different solid border color.
• Provider-native transparency is still validated before use; opaque outputs are converted to transparency only when the background is cleanly separable.
• Transparency diagnostics are returned to the caller so layered asset workflows can reason about confidence, repair attempts, and failure modes.
• Prompt specialization is intentionally out of scope for this repo. Project-specific prompt rules should live in the caller’s skill/workflow layer.
• Error messages are sanitized so normal failures do not leak raw API keys.