GenCurator — Generative AI Model Discovery & Curation

An MCP server built for creative professionals who need to find the right generative AI model for the task at hand — fast and without wading through benchmarks. Ask it in plain language what you're trying to make, and it tells you what to use.

"What's the best model for photorealistic product shots?" "Which image model handles retro illustration best?" "Fastest TTS for my podcast, under $5 per million characters?" "Compare Flux, DALL-E 3, and Midjourney for e-commerce work."

GenCurator is fully platform agnostic. It speaks the open Model Context Protocol standard, so it works with any MCP-compatible client — Claude Desktop, Claude Code, Cursor, Continue, Zed, Cline, Goose, LibreChat, and more. You are never locked in to a single AI provider or app.

If you want to stay truly independent — running models via API keys or locally — Msty is worth a look. It supports MCP, works with any API-based or local model, and is a natural home for GenCurator if you want to test and compare models without committing to one platform.

Contributions are very welcome. See CONTRIBUTING.md — especially if you work in a creative field and have ideas for better model scoring or new data sources.

Think of it as a consultant, not a report

GenCurator fetches live data and surfaces it through conversation — it does not auto-generate a formatted report. The quality and depth of the output depends on the model you are using and the questions you ask. Treat it as an expert you can interrogate:

• Ask follow-up questions. "Why did you rank that one first?" or "Which of these came from GenCurator's data and which from your own knowledge?"
• Ask for sources. The underlying model may add its own recommendations on top of the live data — asking it to separate the two keeps things honest.
• Be specific. The more context you give about your task, format, platform, and constraints, the more useful the recommendation.

There is no single correct answer to "what is the best model" — context always matters, and GenCurator is a starting point for that conversation, not the end of it.

Who this is for

GenCurator is optimised for creative use cases across all generation modalities:

| You make… | Modality | Example query | |-----------|----------|---------------| | Photography / product visuals | Image | "Photorealistic skin tones and natural lighting for portrait work" | | 16:9 renders / video thumbnails | Image | "Cinematic widescreen image, accurate 16:9 aspect ratio, no cropping" | | Illustrations / concept art | Image | "Illustration style that scales cleanly from sticker to billboard" | | Logos / brand assets | Image | "Minimal logo with clean lines, suitable for vectorisation" | | Photorealistic product shots | Image | "High-fidelity product photography for e-commerce, white background" | | Image upscaling / enhancement | Image | "Upscale and restore a low-res photo to print quality without artefacts" | | Short-form video / social content | Video | "Smooth 4-second product clip for Instagram, no flickering" | | Podcasts / voiceovers / dubbing | Audio | "Natural-sounding TTS with Dutch language support, low latency" | | Background music / sound design | Music | "Ambient background music for a meditation app, royalty-free" | | Long-form writing / copywriting | Text | "Best cost-efficient model for bulk article drafts" |

The ranking_recommend tool takes a plain-language description of your task and a priority (quality, speed, cost, or balanced) and returns ranked recommendations from live data.

Data sources

| Source | Modalities | What it provides | API key required | |--------|-----------|------------------|-----------------| | Artificial Analysis | All (text, image, video, audio, music) | Elo rankings, pricing, latency | Yes (free) | | Hugging Face Hub | All | Open model metadata, downloads, community metrics | Optional (for gated models) | | OpenRouter | Text | 300+ models with live pricing | No | | BenchLM | Text | Capability scores by category (coding, reasoning, agentic, …) | No | | LMArena | Text | Human-preference Elo from head-to-head battles across 350+ models (via HuggingFace) | No | | EQ-Bench | Text | LLM-judged creative writing (short + long-form) and emotional-intelligence benchmarks | No |

If a key is missing or a source errors out, GenCurator skips that source, returns what it could gather from the others, and surfaces the skipped sources as warnings in the response.

New to this? Start here

GenCurator is a small program that runs on your computer and gives your AI assistant (Claude, Cursor, etc.) the ability to look up and compare AI models. You connect it once and then just ask questions in plain language.

Here is everything you need to do, step by step.

Step 1 — Check if Node.js is installed

Node.js is the engine that runs GenCurator. Open your terminal (on macOS: press Cmd+Space, type Terminal, press Enter) and run:

node --version

If you see something like v20.11.0, you're good. If you get "command not found", download and install Node.js from nodejs.org — pick the LTS version.

Step 2 — Get a free API key

GenCurator pulls live rankings from Artificial Analysis. Their data is free — create an account and generate an API key at https://artificialanalysis.ai/. It takes about 2 minutes.

Step 3 — Connect to your AI client

GenCurator works with any MCP-compatible app. Two good starting points:

• Msty — if you want to stay platform-independent and switch between API-backed or local models freely. Go to Settings → Model Context Protocol → Add MCP Server, enter gencurator as the name, and paste the JSON config shown in the Msty section below.
• Claude Desktop — open ~/Library/Application Support/Claude/claude_desktop_config.json and add inside "mcpServers":

"gencurator": {
  "command": "npx",
  "args": ["-y", "gencurator-mcp"],
  "env": {
    "ARTIFICIAL_ANALYSIS_API_KEY": "your_key_here"
  }
}

Fully quit and relaunch your client. You're done.

That's it. No downloading, no building. npx fetches GenCurator automatically and always keeps it up to date.

Try it out

Ask Claude: "What's the best image generation model for photorealistic product photography?"

Quick start

1. Get a free API key

GenCurator pulls live rankings from Artificial Analysis — create a free account and generate a key. It takes 2 minutes.

2. Connect to your MCP client

GenCurator works with any MCP-compatible client — it is not tied to any single AI provider or app. Pick yours below:

Msty is a great home for GenCurator if you want to stay platform-independent: it works with any API-backed or locally-running model, so you can ask GenCurator which model to use and immediately switch to it — all within the same app.

In Msty: Settings → Model Context Protocol → Add MCP Server. Enter gencurator as the name and paste the following JSON:

{
  "command": "npx",
  "args": ["-y", "gencurator-mcp"],
  "env": {
    "ARTIFICIAL_ANALYSIS_API_KEY": "your_key_here"
  }
}

Save and toggle the server on. The tools become available to any chat that has MCP enabled.

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{
  "mcpServers": {
    "gencurator": {
      "command": "npx",
      "args": ["-y", "gencurator-mcp"],
      "env": {
        "ARTIFICIAL_ANALYSIS_API_KEY": "your_key_here"
      }
    }
  }
}

Fully quit and relaunch Claude Desktop (Cmd-Q on macOS — closing the window is not enough).

claude mcp add gencurator \
  --env ARTIFICIAL_ANALYSIS_API_KEY=your_key_here \
  -- npx -y gencurator-mcp

Run /mcp inside Claude Code to confirm it's connected.

Edit ~/.cursor/mcp.json (global) or .cursor/mcp.json in a project:

{
  "mcpServers": {
    "gencurator": {
      "command": "npx",
      "args": ["-y", "gencurator-mcp"],
      "env": { "ARTIFICIAL_ANALYSIS_API_KEY": "your_key_here" }
    }
  }
}

Restart Cursor or toggle the server on under Settings → MCP.

Add to your Zed settings.json:

{
  "context_servers": {
    "gencurator": {
      "command": {
        "path": "npx",
        "args": ["-y", "gencurator-mcp"],
        "env": { "ARTIFICIAL_ANALYSIS_API_KEY": "your_key_here" }
      }
    }
  }
}

Most MCP clients accept the same shape: a command, args, and env. Use:

• Command: npx
• Args: ["-y", "gencurator-mcp"]
• Env: ARTIFICIAL_ANALYSIS_API_KEY=your_key_here (and optionally HF_TOKEN)

4. Or run as a standalone HTTP server

TRANSPORT=http PORT=3000 node dist/index.js

The server accepts MCP requests at http://localhost:3000/mcp. Use this when running on a remote machine or when your client prefers a URL over spawning a subprocess.

Tools

ranking_get_leaderboard

Ranked list for any modality. Supports filtering by BenchLM capability category for text.

"Top 10 image generation models right now" "Best text models for agentic tasks" → source=benchlm, category=agentic "Top video models ranked by quality" → modality=video, source=artificial_analysis

ranking_search_models

Cross-source search by name, creator, or keyword. Deduplicates and prefers entries with score data.

"Find all Flux variants" → query=flux, modality=image "Stable Diffusion models for image generation" → query=stable-diffusion, modality=image "Suno or Udio music models" → query=suno, modality=music

ranking_recommend

Plain-language task description + priority → ranked recommendations. The most useful tool for creative workflows.

Image generation:

"16:9 cinematic render, no subject cropping, accurate aspect ratio" → modality=image, priority=quality "Illustration style that works at sticker size and billboard scale" → modality=image, priority=quality "Logo with clean geometric shapes, suitable for vectorisation in Illustrator" → modality=image, priority=quality "Photorealistic product photography, white background, e-commerce" → modality=image, priority=quality "Upscale and restore a degraded scan without introducing artefacts" → modality=image, priority=quality "Concept art for a sci-fi game environment, painterly style" → modality=image, priority=quality "Portrait with accurate skin tones under studio lighting" → modality=image, priority=quality

Video:

"4-second product loop for Instagram, no flickering between frames" → modality=video, priority=quality "Slow-motion nature footage, high temporal consistency" → modality=video, priority=quality

Audio / music:

"Natural-sounding voiceover in Dutch, low latency for live use" → modality=audio, priority=speed "Ambient background track for a meditation app, loops seamlessly" → modality=music, priority=balanced

Text:

"Structured legal document drafting, high accuracy" → modality=text, priority=quality "Cheapest model that still handles multi-step reasoning" → modality=text, priority=cost

ranking_compare

Side-by-side table of 2–5 models. Useful when you've narrowed it down and need to make a final call.

"Compare Flux 1.1 Pro, DALL-E 3, and Ideogram for logo work" "Compare GPT-4o, Claude Sonnet, and Gemini Pro for reasoning tasks"

ranking_cache_status

Check or clear the 1-hour data cache.

Design notes

Accuracy vs. token efficiency. GenCurator deliberately balances result quality against context window cost. Leaderboard output uses compact tables with adaptive columns (only columns that have data are shown). Recommendation output is a table with scores, pricing, VI, and source links. JSON format is available for all tools when you need the full data. This keeps the server useful even in long agentic sessions where context is at a premium.

Graceful degradation. If an API key is missing or a source is unreachable, the server returns whatever it can from other sources and reports the skipped ones as warnings — it never fails the entire request because one source is down.

The AI model layer adds its own knowledge — and its own opinions. GenCurator is a data tool, but it is always used through an AI model (Claude, GPT-4o, Gemini, etc.). That model may supplement the tool data with models or facts from its own training — including models that are not in GenCurator's data sources. This can be useful, but it means the response you see is a mix of live ranked data and the model's own knowledge, which may be outdated or biased toward well-known names.

To get the most reliable output:

• Ask for provenance. After receiving a recommendation, ask: "Which of these came from GenCurator's data and which are from your own knowledge?"
• Ask for links. Request direct URLs to each model's page (e.g. on Hugging Face, OpenRouter, or Artificial Analysis) to verify scores independently.
• Ask clarifying questions. The more specific your task description, the less the model needs to fill gaps with its own assumptions. Include format, style, platform, and constraints upfront.
• Use the GenCurator Bulletin prompt (available in the MCP prompt picker in supported clients) for a structured report that separates data from commentary.

Architecture

src/
├── index.ts              # MCP server setup + tool registration
├── types.ts              # TypeScript interfaces
├── constants.ts          # API URLs, config
├── schemas/index.ts      # Zod input schemas
├── services/
│   ├── api-client.ts     # Generic HTTP client with timeout
│   ├── cache.ts          # In-memory TTL cache
│   ├── artificial-analysis.ts
│   ├── huggingface.ts
│   ├── openrouter.ts
│   └── benchlm.ts
└── tools/
    ├── format.ts         # Compact markdown/JSON formatters
    └── handlers.ts       # Tool handler logic + graceful degradation

Contributing

Contributions are very welcome — especially from people working in creative fields. See CONTRIBUTING.md for a full guide. Priority areas: new data sources (Arena.ai, LMArena, VBench), better recommendation scoring for creative tasks, and qualitative model tags ("strong at illustration", "good for photorealism").

License

MIT