imagengen

An MCP server that generates and edits images via Gemini (Nano Banana), Grok Image, and GPT-image, for MCP clients — like Claude Code — that have no native image generation capability.

• Model lists are discovered live from each provider's API, not hardcoded, so new models show up automatically.

• Defaults to each provider's newest non-top-tier model and a non-maximum quality setting, so a plain "generate an image" request doesn't silently pick the most expensive option.

• Images are saved to disk; tools return the file path.

Tools

Related MCP server: OpenAI Image Generation MCP Server

Providers & API keys

Set only the keys for the providers you want to use. If none are set, the tools return a clear no_provider_configured error.

Choosing a default provider

• If exactly one API key is set, it's used automatically.

• If more than one is set, set IMAGE_PROVIDER_DEFAULT to gemini, grok, or gpt-image to avoid being asked every time.

• If more than one key is set and IMAGE_PROVIDER_DEFAULT is not set, the tools return a needs_provider_choice response — Claude Code is instructed (via the tool descriptions) to check its memory for a previously stated preference, or otherwise ask you, then retry with the provider argument.

Other configuration

Install

npx -y imagengen

Add it to your MCP client:

Claude Code

claude mcp add imagengen \
  -e GEMINI_API_KEY=your-gemini-key \
  -e XAI_API_KEY=your-xai-key \
  -e OPENAI_API_KEY=your-openai-key \
  -e IMAGE_PROVIDER_DEFAULT=gemini \
  -- npx -y imagengen

(Omit any -e you don't need. IMAGE_PROVIDER_DEFAULT is optional — see above.)

Claude Desktop / other MCP clients

Add to your MCP config file (e.g. claude_desktop_config.json):

{
  "mcpServers": {
    "imagengen": {
      "command": "npx",
      "args": ["-y", "imagengen"],
      "env": {
        "GEMINI_API_KEY": "your-gemini-key",
        "XAI_API_KEY": "your-xai-key",
        "OPENAI_API_KEY": "your-openai-key",
        "IMAGE_PROVIDER_DEFAULT": "gemini"
      }
    }
  }
}

Known limitations

• Grok image edits currently support one input image per call (the documented request shape takes a single image field).

• The Gemini provider talks to Google's newer "Interactions" image API (/v1beta/interactions); if Google adjusts that response shape, src/providers/gemini.ts may need a small update.

License

MIT