Labnana

npm version npm downloads CI node License

An MCP server for the Labnana OpenAPI. It enables Claude, Claude Code, and other MCP clients to generate and edit images with Labnana.

Current OpenAPI models (product names mapped to the model parameter):

Product name	model	Provider	Resolutions	Best for
Nano Banana Pro	gemini-3-pro-image	Google	1K / 2K / 4K	In-image text, character consistency, high-fidelity output
Nano Banana 2	gemini-3.1-flash-image	Google	1K / 2K / 4K	Fast iteration and extreme aspect ratios
GPT-Image-2	gpt-image-2	OpenAI	1K / 2K / 4K	Spec-driven generation, layout control, illustration
Wan2.7 Image Pro	wan2.7-image-pro	Alibaba	1K / 2K / 4K¹	Photoreal and poster-style images
Wan2.7 Image	wan2.7-image	Alibaba	1K / 2K	Lower-cost everyday generation
Seedream 5.0 Pro	seedream-5-0-pro	ByteDance	1K / 2K	Coordinate-driven region editing and Chinese instructions

¹ wan2.7-image-pro supports 4K only for text-to-image; generations with reference images are limited to 2K. Treat the Labnana OpenAPI guide as authoritative for model IDs.

GitHub · npm · 中文文档

Installation

Requires Node.js 20.9 or later.

Claude Code

1. Create an API key in the Labnana API Keys console.
2. Make LABNANA_API_KEY available in the environment used by Claude Code.
3. Add the server:

claude mcp add labnana -- npx -y @exoticknight/labnana-mcp

The package can also be started directly with:

npx -y @exoticknight/labnana-mcp

DeepSeek Harness (DSH)

Add the server through DSH's official MCP client plugin. Pin 2.1.0 when you want the version shown by initialize to identify this MCP Apps-capable build:

- id: mcp-labnana
  name: '@deepseek-ai/dsh-mcp-client'
  config:
    serverName: labnana
    transport: stdio
    command: npx
    args: ['-y', '@exoticknight/[email protected]']
    env:
      LABNANA_API_KEY: !!js process.env.LABNANA_API_KEY

DSH's MCP bridge can project supported ImageContent into the calling vision model. Version 2.1 also publishes a standard MCP Apps single-file View for generate_image and get_generation_task. An MCP Apps-capable DSH Web host renders the preview inline; a stock generic DSH tool card may still show the JSON fallback even though the model received the image. This is a client presentation limitation, not a lost generation result.

For a local source checkout, use the same row with command: node, an absolute args path to dist/index.js, and cwd set to this repository. Current stock DSH builds bridge MCP tools but do not consume MCP resources in the generic tool card. Without an MCP Apps host, model vision still works and the card falls back to JSON.

Cursor and VS Code

One-click install (replace the placeholder API key after installing):

Install in Cursor Install in VS Code

Claude Desktop and other MCP clients

Use an equivalent mcpServers configuration:

{
  "mcpServers": {
    "labnana": {
      "command": "npx",
      "args": ["-y", "@exoticknight/labnana-mcp"],
      "env": {
        "LABNANA_API_KEY": "lh_xxxxxxxxx"
      }
    }
  }
}

Local source checkout

npm install
npm run build

Then configure the server with the generated dist/index.js:

claude mcp add labnana -- node <path-to-repo>/dist/index.js

On Windows, use an absolute path such as:

claude mcp add labnana -- node C:/path/to/labnana-mcp/dist/index.js

Configuration

The server reads the following environment variable:

Variable	Required	Description
LABNANA_API_KEY	Yes	Labnana API key.
LABNANA_OUTPUT_DIR	No	Default directory for saved images. Falls back to labnana-images/ under the working directory.

The command-line options below are also supported:

Option	Description
--api-key <key>	Provide the API key for a local process.
--base-url <url>	Override the default API endpoint, https://api.labnana.com.
--output-dir <dir>	Default directory for saved images.

Environment variables are recommended because command-line arguments may be visible in the local process list.

Tools

Tool	Description
generate_image	One-stop text-to-image / image-to-image / editing. Saves the original and returns a bounded MCP image preview, structured metadata, and JSON fallback by default; 4K requests poll internally.
estimate_credits	Estimate the credits required for a generation without generating an image.
get_subscription	Get subscription status, credit balances, and free usage information.
list_generation_tasks	List generation task history with pagination and optional status filtering.
get_generation_task	Get task details and public image URLs (useful after a generate_image timeout).

Usage

Generate an image

{
  "name": "generate_image",
  "arguments": {
    "model": "gemini-3-pro-image",
    "prompt": "Change the background of the image to the grasslands of Inner Mongolia",
    "referenceImages": [
      {
        "fileData": {
          "fileUri": "https://cdn.labnana.com/xxx.png",
          "mimeType": "image/png"
        }
      }
    ],
    "imageConfig": {
      "imageSize": "2K",
      "aspectRatio": "16:9"
    }
  }
}

The default is outputMode=hybrid: the full original is saved to disk while a bounded MCP image preview, structuredContent, and equivalent JSON text are returned together. This provides progressive compatibility across Codex, Claude Code, Claude Desktop, and other MCP clients. Use saveDir to choose the target directory.

All 1K, 2K, 4K, and get_generation_task results use the same envelope:

{
  "schemaVersion": 1,
  "status": "succeeded",
  "taskId": "task-123",
  "images": [
    {
      "index": 0,
      "mimeType": "image/png",
      "width": 4096,
      "height": 4096,
      "byteLength": 18442231,
      "sha256": "...",
      "url": "https://.../original.png",
      "filePath": "C:\\...\\labnana.png",
      "preview": { "included": true, "mimeType": "image/jpeg", "width": 1600, "height": 1600 }
    }
  ]
}

Image base64 appears only in standard MCP ImageContent, never duplicated into text or structured metadata. Full 4K originals are not inlined; previews have a maximum 1600-pixel edge and target a 2 MiB byte ceiling, while filePath/url locate the original.

Use referenceImages for image-to-image generation and editing. fileData.fileUri supports gs:// and https://; small images can be passed as base64 through inlineData.data.

4K requests automatically run as asynchronous tasks: the server creates the task, polls with rate-limit backoff, downloads and saves originals, and creates previews. If the wait exceeds timeoutSeconds (default 300), the result has status=pending and a taskId; this is not treated as generation failure. Fetch the final preview and original URL later with get_generation_task.

Credit estimation

{
  "name": "estimate_credits",
  "arguments": {
    "prompt": "A Shiba Inu running through a snowy landscape",
    "imageConfig": {
      "imageSize": "4K"
    }
  }
}

Parameters

• model (default: gemini-3-pro-image): gemini-3-pro-image, gemini-3.1-flash-image, gpt-image-2, wan2.7-image-pro, wan2.7-image, or seedream-5-0-pro. The provider is derived from the model automatically.
• imageConfig.imageSize: 1K, 2K, or 4K. wan2.7-image and seedream-5-0-pro do not support 4K; wan2.7-image-pro also disallows 4K when reference images are present.
• imageConfig.aspectRatio: 1:1, 2:3, 3:2, 3:4, 4:3, 9:16, 16:9, 21:9, 1:4, 4:1, 1:8, or 8:1. GPT-Image-2 may omit this field and let the service choose; Wan2.7 supports only 1:1, 16:9, 9:16, 4:3, and 3:4.
• referenceImages: OpenAPI allows up to 14 for Gemini, 4 for GPT-Image-2, 9 for Wan2.7, and 10 for Seedream. These are API limits, not the separate upload limits of the web generator.
• Seedream precise editing: put the source image in referenceImages and describe the target region and change in prompt using absolute coordinates from the top-left origin. OpenAPI has no separate mask or region parameter.
• outputMode (generate_image only): hybrid (default, save originals and inline bounded previews), file (save originals and return metadata only), or inline (do not save to saveDir; return a bounded preview, and persist the original to the default recovery directory only when the upstream response has no original URL).
• saveDir / timeoutSeconds (generate_image only): target directory for hybrid/file mode, and the maximum wait for async (4K) generations.

Credit summary

Model	1K	2K	4K
gemini-3-pro-image	15	15	30
gemini-3.1-flash-image	10	10	20
gpt-image-2	4	6	10
wan2.7-image-pro	6	8	12 (text-to-image only)
wan2.7-image	4	6	Not supported
seedream-5-0-pro	6	15	Not supported

Errors

API errors are returned as { code, message } and exposed as MCP results with isError: true.

Code	Meaning	Recommendation
21007	Invalid API key	Check LABNANA_API_KEY.
26004	Insufficient credits	Check the subscription or upgrade the plan.
29003	Invalid parameters	Check required fields and model-specific limits.
29998	Too many requests	Retry with a 20–30 second backoff.

Development

npm install
npm run typecheck
npm test

To inspect the real single-file MCP App without calling Labnana or spending credits, run npm run test:ui and open http://127.0.0.1:4173/test/mcp-app-host.html. The local host sends a generated 640×360 PNG fixture through the official AppBridge so the image, status, and metadata can be checked visually.

Links

• Labnana integration guide
• Labnana OpenAPI reference
• Credit guide
• GitHub repository
• Technical support: [email protected]

License

Apache-2.0