Labnana MCP Server

labnana-mcp

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):

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 · MCP Registry · Glama · 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 2.1.1 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/labnana-mcp@2.1.1']
    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):

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:

The command-line options below are also supported:

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

Tools

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

Errors

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

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
• Official MCP Registry
• Glama directory
• Technical support: support@marswave.ai

License

Apache-2.0