mcp-imagenate

An MCP server for image generation using multiple providers: Google Gemini, OpenAI (gpt-image), BFL FLUX, and Reve — plus short video clips through Google Gemini Omni.

Providers & Models

Google Gemini (Nano Banana)

Google Gemini Omni (video)

Uses the same GEMINI_API_KEY. Exposed through a separate generate_video tool — see Tool: generate_video.

OpenAI

These are the only models here that can return a transparent background — see Transparent backgrounds.

BFL FLUX

Reve

This provider calls Reve's v2/image/create endpoint. latest is the only version alias v2 exposes, and it is what the response reports back, so there is no dated build to pin to. Do not confuse it with the v1 endpoints, which still serve the older reve-create@20250915 model.

Things worth knowing before sending Reve a prompt written for another provider:

• resolution is ignored — Reve has no size parameter and returns its own large output. Exact dimensions vary between requests: 16:9 came back as both 5408x3072 and 5376x3072, and 3:4 as 3456x4800.

• Prompts are capped at 4,000 characters, and this provider rejects longer ones before spending a request.

• inputImages become v2 references. Reve accepts at most eight; a longer list is rejected before any of the files are read.

• The saved file's extension follows the format Reve actually returned (PNG, JPEG or WebP), which is detected from the bytes rather than assumed.

• A generation costs 150 credits (about $0.20) and typically takes 40-80 seconds. Give any proxy or job runner in front of it a timeout of at least 120 seconds.

Related MCP server: Imagen Tools MCP Server

Requirements

• Node.js 20+

• At least one provider API key

Installation

npx mcp-imagenate

Or install globally:

npm install -g mcp-imagenate

Setup

Set API keys for the providers you want to use:

# Google Gemini (at least one)
export GEMINI_API_KEY=your_key_here
# or
export NANO_BANANA_API_KEY=your_key_here

# OpenAI (at least one)
export OPENAI_API_KEY=your_key_here
# or
export GPT_IMAGE_API_KEY=your_key_here

# BFL FLUX
export BFL_API_KEY=your_key_here

# Reve (at least one)
export REVE_API_KEY=your_key_here
# or
export REVE_API_TOKEN=your_key_here

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "mcp-imagenate": {
      "command": "npx",
      "args": ["mcp-imagenate"],
      "env": {
        "GEMINI_API_KEY": "your_key_here",
        "NANO_BANANA_OUTPUT_DIR": "/path/to/image/output"
      }
    }
  }
}

Environment Variables

* At least one provider API key must be set.

Tool: generate_image

Parameters

Supported aspect ratios

1:1, 2:3, 3:2, 3:4, 4:3, 9:16, 16:9, 21:9

Transparent backgrounds

background: "transparent" saves a PNG with an alpha channel, which is useful for cutting out a subject to place on a slide or over another image.

Only the OpenAI gpt-image models can do this. Asking any other model (nano-banana-*, flux-2-*, reve-image) for a transparent background fails with an error rather than quietly returning an opaque image — the request is rejected before it is sent, so nothing is spent on it. Writing "transparent background" into the prompt does not help either: those providers have no transparency mode at all.

"opaque" forces a filled background on every provider that reads the field, and "auto" — the default — leaves the choice to the model, which is what this server has always done.

Response

Returns a JSON object:

{
  "model": "gemini-3.1-flash-image-preview",
  "savedFiles": ["/path/to/image-1.png"],
  "settings": {
    "resolution": "1K",
    "aspectRatio": "9:16",
    "mode": "image",
    "background": "auto"
  },
  "description": "..."
}

description is only present when mode is "image_and_text".

Tool: generate_video

Available when a Google key is configured. Generates one clip with audio and saves it as an mp4.

Parameters

Things worth knowing:

• A single request is capped at 10 s by the model. To go longer, pass the returned interactionId back as previousInteractionId; each extension adds up to 10 s, and the whole clip is returned each time.

• Text in the prompt is rendered on screen as written, including non-Latin scripts, though Google only documents English as fully supported.

• Clips are fetched through Google's file endpoint rather than inlined in the JSON response, as the API documentation recommends above 4 MB. Expect one extra request per generation.

• 720p costs about $0.10 per second of output; there is no free tier for this model.

Response

{
  "model": "gemini-omni-1.1-flash",
  "savedFile": "/path/to/1788347054697-491db547.mp4",
  "settings": {
    "durationSeconds": 5,
    "resolution": "720p",
    "aspectRatio": "16:9"
  },
  "interactionId": "v1_...",
  "description": "..."
}

description is only present when the model returns text alongside the clip.

Use as a library

Besides the standalone MCP server, this package can be embedded in another host — an app, or another MCP server that wants to expose image generation as its own tool.

import { createRegistry, generateImageToDisk } from "mcp-imagenate";

// Keys are passed in explicitly; nothing here reads process.env.
const registry = createRegistry({ openai: myOpenAIKey, google: myGoogleKey });

if (registry.models.length === 0) {
  throw new Error("No image provider is configured");
}

const outcome = await generateImageToDisk({
  registry,
  prompt: "a calico cat asleep on a warm keyboard",
  model: registry.defaultModel!,
  aspectRatio: "16:9",
  outputDir: "/somewhere/to/write",
  // outputBaseDir defaults to null, meaning no path sandboxing. Set it to a
  // directory to confine both output and input paths within that directory.
});

console.log(outcome.savedFiles);

generateImageToDisk takes the same options as the tool, so background: "transparent" throws for a model that cannot deliver an alpha channel. Check registry.resolve(model).supportsTransparentBackground first if the model is not one you chose yourself.

The library entry point never reads process.env, writes to stdio, or exits the process. To read keys from the conventional environment variables anyway, use the keysFromEnv() helper. The standalone server is available at mcp-imagenate/server.

Security

• Path sandboxing: When NANO_BANANA_OUTPUT_DIR is set, both output and input image paths are sandboxed within this directory. Symlinks that resolve outside the sandbox are rejected. For library embedders this is opt-in via outputBaseDir, since the host usually controls which paths reach the call.

• Input validation: Input images are validated for format (PNG/JPEG/WEBP/GIF) and size (max 20 MB). Video durations outside the model's range are rejected before any request is sent.

• API key validation: The server exits immediately if no API keys are configured. The library reports this as an empty registry instead, leaving the decision to the host.

License

MIT