openrouter-imgen-mcp

Give your AI assistant the ability to generate and edit images. One install, multiple models, no vendor lock-in.

Screenshots

| Before | After | |--------|-------| | | |

Works with Claude Code, Claude Desktop, Cursor, VS Code, Windsurf, and any MCP client.

Why This

• One API key, every image model. Gemini, GPT-5, and more through OpenRouter. Switch models with a word, not a code change.
• Generate, edit, and batch. Text-to-image, image editing with references, and batch operations (up to 20 parallel tasks).
• Style presets. Say "in watercolor style" or "as pixel art" and get consistent results. 8 built-in styles.
• Variations. Generate up to 4 takes from one prompt. Pick the best one.
• Smart file output. Save to a path or a directory. Files get descriptive names automatically.
• Pay for what you use. OpenRouter's per-token and per-image pricing means no subscriptions.

How to Install

Get an OpenRouter API key, then add the server to your client.

Claude Code

# macOS / Linux
claude mcp add openrouter-imgen-mcp -s user -e OPENROUTER_API_KEY=YOURAPIKEY -- npx -y openrouter-imgen-mcp@latest

# Windows
claude mcp add openrouter-imgen-mcp -s user -e OPENROUTER_API_KEY=YOURAPIKEY -- cmd /c npx -y openrouter-imgen-mcp@latest

Claude Desktop / Cursor / VS Code / Windsurf

Add to your MCP config JSON:

{
  "mcpServers": {
    "openrouter-imgen-mcp": {
      "command": "npx",
      "args": ["-y", "openrouter-imgen-mcp@latest"],
      "env": {
        "OPENROUTER_API_KEY": "YOURAPIKEY"
      }
    }
  }
}

What You Can Do

Simply chat with your assistant:

Generate a product photo of a ceramic mug on a wooden table, warm lighting

Generate 3 variations of a logo concept for a coffee brand, pixel-art style

Edit /path/to/headshot.png: make the background a gradient, keep the subject

Batch generate 5 social media banners: summer sale, back to school, holiday promo,
new arrivals, free shipping. All in illustration style, 16:9, save to ./banners/

List image models. What's cheapest for quick drafts?

Quality Presets

Skip the model IDs. Just say "fast", "balanced", or "quality":

| Preset | Model | When to use | |--------|-------|-------------| | fast | Gemini 2.5 Flash | Drafts, iteration, keeping costs low | | balanced | Gemini 3 Pro | Day-to-day use (default) | | quality | GPT-5 Image | Final deliverables, max fidelity |

Or pass any OpenRouter model ID directly:

Generate a cat using model google/gemini-2.5-flash-image

Style Presets

Keywords modifies the style:

| Style | What you get | |-------|-------------| | photo | Photorealistic, DSLR, natural lighting | | illustration | Clean digital art, vibrant colors | | watercolor | Soft washes, paper texture | | pixel-art | Retro 16-bit, limited palette | | 3d-render | Studio-lit, physically based rendering | | anime | Cel-shaded, Japanese animation | | sketch | Pencil on paper, monochrome | | oil-painting | Brushstrokes, impasto, classical |

Tools Reference

Convenient tools:

| Tool | What it does | |------|-------------| | generate_image | Text-to-image with quality/style presets, variations (1-4), aspect ratio, file output | | edit_image | Transform images with a prompt + up to 12 reference images (files, URLs, data URIs) | | batch_images | Run up to 20 generate/edit tasks in parallel with concurrency control | | list_image_models | Browse available models with pricing | | check_balance | Check your OpenRouter credit balance |

Supported output: 1:1, 2:3, 3:2, 16:9, 9:16, 4:3, 3:4 aspect ratios. 0.5K, 1K, 2K, 4K sizes. PNG, JPEG, WebP, GIF inputs.

Headless Configuration

| Flag | Env Var | Default | |------|---------|---------| | --api-key | OPENROUTER_API_KEY | (required) | | --default-model | OPENROUTER_IMG_DEFAULT_MODEL | google/gemini-3-pro-image-preview | | --batch-concurrency | OPENROUTER_IMG_BATCH_CONCURRENCY | 3 | | --cache-ttl | OPENROUTER_IMG_CACHE_TTL | 300000 (5 min) | | --log-level | OPENROUTER_IMG_LOG_LEVEL | warn |

Security

API keys never appear in logs or error messages. File paths block directory traversal. Image URLs are validated against SSRF patterns. Batch operations are memory-capped at 200MB. All inputs are schema-validated.

For Developers

To build for devs:

pnpm install && pnpm build && pnpm test  # 57 tests

License

MIT

Made with <3 at Bitcoin.com