Delphi MCP & Agentic AI Training: Make Your ERP Agentic
Two hands-on days to go from the MCP protocol to an AI agent embedded in your Delphi ERP, with mcp-server-delphi and DMVCFramework. In Italian, English or Spanish, on-site or remote.🇮🇹 Italiano • 🇩🇪 Deutsch • 🇪🇸 Español • 🇧🇷 Português
The hands-on course that brings an AI agent inside your Delphi ERP: from the MCP protocol to the embedded agentic loop, with mcp-server-delphi and DMVCFramework.
There’s a moment, in every paradigm shift, when a technology stops being “something interesting to watch” and becomes “something I need to know how to do”. For agentic AI in the Delphi world, that moment is now.
I’m not talking about chatbots. I’m not talking about pasting a ChatGPT answer into a TMemo. I’m talking about an AI agent that lives inside your application, one that understands a request in natural language, orchestrates your tools, follows your business rules and stops to ask for your confirmation when needed. Written in Delphi. With your data never leaving your infrastructure.
Until recently, for a Delphi developer this was a distant frontier. Today it’s a two-day course.
Scenarios from the near future
Monday morning. It's 8:47 on a Monday. The sales manager opens the ERP and instead of seven screens finds a sentence: "12 orders came in over the weekend, two above 50k. Customer Rossi has exceeded their credit limit: the order is waiting for your decision. I've prepared the summary and a draft reply." No dashboard to interpret: the ERP has already read, cross-checked and prepared everything. What's left for him is the part that matters: deciding.
Why mcp-server-delphi Changed the Game in the Delphi World
When the Model Context Protocol arrived, the standard LLMs use to talk to external software, the natural first step was: “let’s expose the ERP’s tools as an MCP server, so Claude or Gemini can use them”. Useful, but it’s half the picture. In that scenario the intelligence lives outside, in someone else’s client, and your application remains a passive supplier of functions.
And this needs to be said clearly: stopping at a lone MCP server, for many real business scenarios, is limiting and delivers little value. It means your users have to leave the ERP and work inside Claude or Gemini; that you control neither the user experience nor the moment when the AI must stop and ask for confirmation; that your data passes through a third-party client. An MCP server is a great starting point, but on its own it doesn’t transform your software: it merely makes it “drivable from the outside”. In most business cases what you really need is the agent inside the application, not an external remote control.
The open source mcp-server-delphi library (Apache 2.0, integrated with DMVCFramework) was a small revolution in the Delphi world precisely because it doesn’t stop at the server. In a single library you get the three parts you need to build a complete agentic application:
| Component | What it gives you |
|---|---|
| 🟢 MCP Server | expose your ERP’s tools to any AI client (Claude Desktop, Gemini CLI…), stdio and HTTP transports |
| 🔵 MCP Client | your Delphi app connects to other MCP servers and consumes their tools |
| 🟣 Agentic loop | the heart: an agent written in Delphi that reasons, calls tools in sequence and gets the task done |
It’s this third part that makes the difference. With the embedded agentic loop the intelligence runs inside your .exe, inherits the session and permissions of the already logged-in user, and calls Pascal code you wrote. Your application doesn’t “support MCP”: it becomes agentic.
Scenarios from the near future
Conversational management accounting. "Why did March's margin drop by two points?" You can't ask a database a question like that: you ask a database for numbers, and you build the "why" yourself with half a day of extractions, pivots and phone calls. You can ask an agent for the "why" directly: it digs through jobs, price lists and purchase costs, forms hypotheses, checks them against the data, and answers: "Three jobs closed below margin, all from the same salesperson, all with the extra discount applied by hand. Want the details?" The data was already there. What was missing was someone to go look for it while you do something else.
And there’s a detail that matters more than it seems: if you already have RESTful APIs built with DMVCFramework in production, mcp-server-delphi turns them into agentic tools without rewriting them. With a single line your existing REST engine gets exposed as an MCP server, and every endpoint becomes a function the agent can call. Years of already-written APIs suddenly become agent capabilities.
This shortcut exists for a precise reason: mcp-server-delphi isn’t an isolated library, it’s an extension of DMVCFramework, the most popular Delphi web framework on GitHub. It already knows your controllers, your routes and their documentation, and hands them to the AI at almost zero cost. No other approach in the Delphi world starts from such a solid foundation.
Why Companies Want This (and Why You’ll Want It Too)
The numbers speak for themselves: between June and July alone, more than 20 European companies developing in Delphi asked us for consulting to understand what’s concretely possible with this library and how to bring it into their applications. Companies with software in production that want to know where to start, not academic curiosity.
In recent weeks the question I get most often is no longer “what is MCP?”, but “how do I put it into my ERP?”. Companies have understood the important thing: they already have the value. The data is there. The business logic, built up over years of development, is there. The rules are there.
Only one piece is missing: an agent that knows how to use all of it on command, in natural language.
Picture a sales director who, instead of opening five screens, types “handle this order email”, and the agent reads the request, identifies the customer, infers from the history the “usual discount” never written in any field, prepares the draft order and the confirmation email, and stops to wait for the human go-ahead. That’s exactly the pattern you’ll learn to build during the course.
Scenarios from the near future
The payment schedule that (almost) writes itself. "Prepare the reminders for invoices more than 30 days overdue, but use a soft tone with long-standing customers." The agent scans the payment schedule, recognizes who always pays a few days late and who is instead becoming a problem, prepares different emails for different cases, and puts them all in the approval queue. Twenty minutes of admin work become two minutes of review.
Whoever brings this capability into their application first gives their users much more than “one more AI feature”: they give them a completely new way of working with the software. It’s a competitive advantage, and it’s within reach of anyone who knows Delphi like you do.
The Idea Behind the Course: We Build an MCP Server, Then Evolve It Into an Agent
This isn’t a theory course. It’s a hands-on path where, starting from zero, we build a single application that grows with us. Every module adds a piece, and you can compare it with the previous one to see exactly what changed.
The trajectory is precise and deliberately progressive:
- First we understand the AI: what an LLM does (and what it doesn’t do), and why it needs tools.
- Then we build a minimal MCP server: a single tool, in ~90 lines, connected to a real AI client.
- We enrich it: typed and optional parameters, all the result types (numbers, objects, datasets, media), the
FromDataSet“golden bridge” that turns a FireDAC query into JSON the model understands. - We expose it over HTTP: and face the first real architectural decision.
- We flip the perspective, from server to agent: the same server, now driven by an agentic loop running inside the Delphi app.
- We secure it: because an agent with write tools is the part that decides whether the project actually goes to production.
- We take it home: from the real-world case to your ERP.
At the end of the two days you won’t have watched an agent work: you’ll have built one, and you’ll know how to build it again on your own software.
Scenarios from the near future
The warehouse that notices things. Today's warehouse tells you how many units you have. The agentic one notices that item X is dropping faster than usual, that the regular supplier has just stretched their delivery times, and that in six weeks there's the seasonal spike that catches you off guard every year. And it puts a draft purchase proposal in front of you, with the reasoning spelled out in black and white. You read it, correct it, approve it. The experience stays yours; the grind of holding all the signals together doesn't.
The Program (Two Days)
Day 1: From the Protocol to the MCP Server
- M0 · Opening demo and map: where we’re headed, shown before it’s even explained.
- M1 · Agentic AI in half an hour: LLMs, tool use and the agentic loop, explained for programmers.
- M2 · The MCP protocol: Host, Client, Server; tools, resources, prompts; why Delphi is a perfect fit (RTTI + attributes → schema generated on its own).
- M3 · Your first MCP server 🧪: the first tool, auto-registration, stdio discipline, hooking it up to a real AI client.
- M4 · A rich server 🧪: parameters, result types and
FromDataSet, the way a Delphi ERP talks to the AI. - M5 · HTTP and the architectural decision: dual transport and the choices that follow from it.
Day 2: From the Agent to Production
- M6 · From server to agent 🧪: hosting the model, configuring the endpoint, running the embedded agentic loop.
- M7 · Trust & Security: GDPR, prompt injection, governance: read-only first, writes with human confirmation; the man-in-the-loop rule.
- M8 · Reliability & trust engineering: the real-world
mcp-firebirdcase: Guard pattern, audit, deterministic evals. A production MCP server gets tested like real software. - M9 · AgenticERP deconstructed: the full demo: a single
.exethat is both agent + MCP server, built on the same patterns as the course, at enterprise scale. - M10 · Your first agent on your own ERP: the method to apply all of this to your application, not to a sample.
- M11 · Production & beyond + Q&A: costs,
max_turns, REST→MCP bridge, deployment, and your questions.
Every 🧪 is a lab: you write code, you compile, you watch the model call the tools you just created.
Scenarios from the near future
The senior colleague who never goes on vacation. Every company has that one person who knows how to issue the credit note for a partial return with split-payment VAT. Then she goes on vacation, or retires, and that knowledge walks out the door with her. An agent inside the ERP is that senior colleague, always available: you ask how it's done, it explains and meanwhile prepares the screen already filled in. Tribal knowledge stops being tribal.
In Italian, English or Spanish. On-Site or Remote.
The course is built to be flexible:
- 🗣️ Three languages: it can be delivered in Italian, English or Spanish.
- 🏢 On-site: right at your company, with examples tailored to your domain.
- 💻 Remote: the same hands-on experience, wherever your team is.
The course fits every Delphi developer. If you’ve never touched LLMs, agents or MCP, you start from the foundations, with nothing taken for granted. If you’ve already experimented, you get a structured path that lines up the pieces and goes all the way to an agent in production.
The long view
Let's look a little further ahead. The ERP of the near future won't just have users: it will have agents that prepare the next day's drafts overnight, that answer questions instead of the manual, that notice what the human eye misses. But with a crystal-clear hierarchy: they propose, they don't decide. Every write goes through a human confirmation, every action leaves a trace in the audit log. The paradox is only apparent: the more capable the agent, the more your job goes back to being what it always was, understanding the customer and deciding well. The rest is digital grunt work.
We’re Putting Together a Class: Get On Board
We’re assembling a new class for MCP & Agentic AI with Delphi. If you want to be among the first to bring an AI agent inside your ERP (and do it with method, safely, with code you understand all the way down), this is the moment and this is the starting point.
- 📩 For registrations, dates and formats (single-company or multi-company): write to professionals@bittime.it. The dedicated course page is coming soon; in the meantime, email is the direct channel.
- 💬 Already have a project in mind? Beyond training, we offer specialized consulting to concretely integrate agentic AI into existing systems (Delphi, Python, C# and more): contact us.
Did you attend the "Architetture AI per LLM: dalle fondamenta tecniche all'operatività aziendale" course (AI Architectures for LLMs)? You're entitled to a 20% discount on this new course. If you're interested in the course (and the discount), write to professionals@bittime.it.
Your Delphi ERP already has all the value it needs. The only thing missing is an agent living inside it. In two days, you learn how to build it.
Frequently Asked Questions
In which languages is the MCP and Agentic AI with Delphi training delivered? The course can be delivered in Italian, English or Spanish.
Is it held on-site or remotely? Both: on-site at your company or remotely, with the same hands-on approach and the same labs.
How long is it? Two days. Day one takes you from the MCP protocol to a rich MCP server; day two flips the perspective, from server to embedded AI agent, and covers security and production.
Who is it for? Delphi developers, including those with no prior experience with LLMs, agents or MCP. We start from the foundations and take nothing for granted.
What are the prerequisites? Knowing Delphi. Familiarity with DMVCFramework helps, but it’s not required: the examples start from scratch.
Can I reuse my existing DMVCFramework REST APIs?
Yes. Since mcp-server-delphi is an extension of DMVCFramework, a REST engine already in production can be exposed as an MCP server with a single line, turning your endpoints into agentic tools without rewriting them. In the course we show how to do it and how to refine them for AI use.
What do you actually build during the course? A single application that grows module after module: from a minimal MCP server to an embedded AI agent that orchestrates your tools, follows your business rules and stops to ask for confirmation when needed.
How does the 20% discount work? If you’ve already attended the “Architetture AI per LLM: dalle fondamenta tecniche all’operatività aziendale” course (AI Architectures for LLMs), you’re entitled to a 20% discount on this course. To claim it, write to professionals@bittime.it.
How do I register? Write to professionals@bittime.it. The dedicated course page is coming soon.
Comments
comments powered by Disqus