Hey Stefano, 

have you tried asking the AI the best way to ask it questions?

Dears,

After following this interesting discussion about AI and PICK/MultiValue code, I would like to share an approach that I have developed which goes far beyond simply generating BASIC code with ChatGPT or Gemini. I want to clarify that this approach is specifically designed for analyzing information from D3 databases using natural language.

The MCP-Pick Revolution: Connecting Legacy Systems with Conversational AI

I have developed a bidirectional bridge between Rocket D3 systems and advanced AI models such as Claude and Gemini, using the Model Context Protocol (MCP). This technology allows our MultiValue systems to speak directly with AI agents through natural language, without the need for migration or rewriting.

How does my solution work?

Structured data exposure: I use standard AQL commands along with mvstoolkit to expose PICK files as HTTP endpoints that return structured JSON:

LIST INVOICE NAMEPROD NAMEBUSNIESS CODE CUSTOMER CODEPROD PRODQUANTITY PRODTOTAL DATEINVOICE

Orchestration with n8n: I implement automated workflows in n8n that invoke these endpoints and send the data to the MCP through Server-Sent Events (SSE).

Bidirectional communication: The AI agent (Claude/Gemini) processes this data and can respond to complex queries such as "What is the total billing by customer in April?" with contextualized analysis.

Tangible Results

The system I have implemented allows:

Natural language queries: Directly ask about data stored in Pick D3

Automated financial analysis: Calculations, monetary formats, and comparative analyses without additional programming

Modern interface generation: Complete dashboards in React/Tailwind automatically generated

Instant Analysis of MultiValue Data

The most revolutionary aspect of my approach is the ability to analyze Pick D3 data instantly and in any conceivable way. In my example, I work with a single billing file and a few fields (NAMEPROD, NAMEBUSNIESS, CODE, CUSTOMER, CODEPROD, PRODQUANTITY, PRODTOTAL), but this is just the tip of the iceberg.

The beauty of the system is that I don't need to predefine reports, dashboards, or specific analyses - I simply ask what I need to know in natural language. "Show me customer purchasing behavior by region," "Analyze the seasonality of our sales by product," or "Identify potential fraud patterns in the last 1000 transactions" - all of this is generated in a matter of minutes, without a single additional line of code. It just takes imagination and the AI handles the rest.

The most impactful example: with a single prompt ("Use MCP PICK to build an interactive financial dashboard in React + Tailwind CSS"), I obtained a complete website with graphs, tables, and risk analysis based directly on my MultiValue data.

Key Difference with Code Generation

Unlike asking AI to generate PICK code (with the limitations mentioned by several of you), my approach:

Leverages the best of both worlds: the stability and robustness of our legacy systems with the flexibility and capabilities of modern AIs

Does not require PICK programming knowledge to obtain modern interfaces and analyses

Allows real-time updating of dashboards and visualizations

This approach mitigates many risks, as the AI does not generate the code that executes critical operations - it only interprets, visualizes, and analyzes data already processed by proven systems.

You can see an example of the generated dashboard here: https://claude.ai/public/artifacts/5d41270f-e6af-4d30-8b08-ca79cc3c4992

I am available to answer questions about this implementation and share more technical details.

In my opinion, this is the true potential of AI. Best regards and I hope this implementation is useful and you can apply it in your applications.

I've attached a video showing how I interact in real-time with Claude desktop and the Pick D3 MCP to generate a portal in less than 5 minutes without writing a single line of code.

Best regards,

Fausto Paredes

https://faustoparedesia.com/

Hi Fausto,

Very interesting post. I was dabbling in MCP and Claude Desktop myself recently, but just for experimental reasons and with Python. I find it fascinating that it is able to use agents to basically do tasks for you. We also have MVIS, so reading your use case was very cool to see.

On the security side of things though, I was wondering that by asking an LLM like Gemini, ChatGPT, or Claude questions about the JSON files you are producing from MVIS, wouldn't it be exposing that data to them? That has been one of the obstacles for us moving forward. In order for us to use our data, we would need the hardware infrastructure to be able to try and host a local AI model, which aren't as powerful as the big LLMs. On the one hand, we see all the great things that these foundational LLMs can do and we want it to be able to analyze our data and do all these things for us. But at the same time, by doing so, we are basically giving them our data which will then be used by them to further train their models. Then we hear news like how, allegedly, Deepseek used OpenAI's training data to help train their own model, which they then released as open source for everyone to use and fine tune. Then at some point, it was found that Deepseek had a vulnerability in a database they were using that potentially exposed their data to anyone who was looking for it at the time. So in those instances, potentially hackers could scrape that data and get a hold of anything in there to analyze and use for their own purposes.

https://www.darkreading.com/cyberattacks-data-breaches/deepseek-breach-opens-floodgates-dark-web

I think right now, is kind of the Wild Wild West of AI innovation. Both good guys and bad guys are trying to get an edge where they can, and unfortunately, sometimes, corners can be cut. I don't think enough is being done in terms of AI security yet. So there are a lot of potential pitfalls we all need to keep an eye out for. Here are a couple of other examples:

Scrutinize the MCP servers you use:

https://invariantlabs.ai/blog/mcp-security-notification-tool-poisoning-attacks

This hacker that stole Disney data, posted an AI image generation tool on Github that a Disney employee used. But the tool had malicious components to it that allowed the hacker to harvest credentials and then gain access to Disney's data:

https://www.securityweek.com/man-admits-hacking-disney-and-leaking-data-disguised-as-hacktivist/

In conclusion, as we all continue to strive forth and charter into unknown AI territory, we still need to protect ourselves while doing so. That is until the Matrix enslaves us all... or Terminators.. or Zombie apocalypse... or for the hopeful ones, a Star Trek future! :)

Now this is a thread I'll enjoy following.

Has anyone had any success with using an AIs to generate MultiValue code ?

Could this help bridge the skills gap with newbies coming in and companies looking at future-proofing their systems.