# Lulu-The-Narwhal/dali-mcp [Health: Active]

**Category:** 🎥 Multimedia Process  
**Repository:** https://github.com/Lulu-The-Narwhal/dali-mcp  
**GitHub Stars:** 8  
**npm Downloads (last month):** 173  
**Views:** 3  
**Installs:** 0  
**Upvotes:** 0  
**Directory Page:** https://allmcps.com/mcp/lulu-the-narwhal-dali-mcp

## Description
Scores AI image/video generation prompts (0-100) against 18 models (Veo 3, Sora, Kling, Runway, Seedance, Midjourney, Flux, and more) before you spend a generation credit, then rewrites weak ones. Free, hosted remote MCP + npm/PyPI packages.

## Claude Desktop Quick Installation
Install path detected from listing signals. Uses `npx` (confidence: high):

```json
"mcpServers": {
  "dali-mcp": {
    "command": "npx",
    "args": ["-y","wrapper"]
  }
}
```

## Documentation

## What Lulu-The-Narwhal/dali-mcp MCP server does

Lulu-The-Narwhal/dali-mcp MCP server provides pre-generation review for AI image and video work, with a focus on advertising creatives. It evaluates both the text prompt and, through the hosted service, the resulting image. Prompt scores range from 0 to 100 and include a grade, dimension-level feedback, and a verdict. Low-scoring prompts can also produce an enhancement brief describing what to change.

The service supports 18 generation models, including Veo 3, Sora, Kling, Runway, Seedance, Midjourney, and Flux. Model-specific scoring considers factors such as camera movement, motion, lighting, and language expected by the selected model.

Creative review is based on a corpus of public ads classified by how long they remained active, rather than on measured conversion rate for each ad. The hosted system can identify visually similar winning ads, highlight attributes that appear more often in those ads, and report apparent generation defects.

## How it works

Lulu-The-Narwhal/dali-mcp MCP server has hosted and local usage modes. The hosted MCP endpoint is the broader option: it provides prompt scoring, image scoring, and access to the winning-ad corpus. A local package is also available, but its documented scope is limited to prompt-scoring tools.

The underlying corpus is assembled from the public Meta Ad Library. Ads are processed with vision analysis, pattern prevalence calculations, and vector search over embedded winner creatives. The service reads prebuilt data stores while serving requests; it does not scrape or embed ads during each tool call.

For an image URL, `score_creative` returns a score based on similarity to proven winners, examples of similar ads, missing winning attributes, and defects. `score_creative_from_view` reviews an image attached or pasted into the conversation without requiring a URL. `analyze_winning_formula` accepts an exported set of ads with performance data and compares the user's winners and losers with an industry median.

## Setup and configuration

The hosted endpoint is `https://dali.getlulu.dev/mcp`. Cursor and Windsurf can connect to it through their MCP configuration using that URL. Stdio-only clients can use the `dali-mcp` npm wrapper with `npx`.

A local installation is available through PyPI:

```bash
pip install dali-mcp
claude mcp add dali -- python -m dali.server
```

The README states that self-hosting does not require authentication. The hosted connection also has no credential configuration documented in the provided material. The local package does not provide the hosted creative-scoring tools or the winning-ad corpus, so use the hosted endpoint when those capabilities are required.

## Tools and capabilities

- `score_prompt(prompt, model, category?)`: grades a prompt, explains the result by dimension, and returns a verdict.
- `enhance_prompt(prompt, model, category?)`: produces a structured brief for rewriting; the calling LLM writes the final enhanced prompt.
- `score_creative(image_url, category)`: scores an image URL against proven winners and reports similarities, missing attributes, and defects.
- `score_creative_from_view(category, …)`: evaluates an image visible in the conversation without a URL.
- `analyze_winning_formula(csv, category, email)`: compares supplied ad performance data with the user's winners, losers, and industry median.

The prompt tools can accept an optional category, allowing feedback to incorporate category-specific winning patterns. The README describes eight covered verticals, including beauty, wellness, supplements, fitness, food, apparel, technology, and pets.

_Full upstream README: https://allmcps.com/mcp/lulu-the-narwhal-dali-mcp/readme_