Integrate applications in tools such as OpenCode, Claude Code or similar with MCP servers.
Current AI tools are extremely powerful. What if the AI could control your own applications, for example to retrieve and process data? All you need to do is implement an MCP server. Two examples from my everyday life:
Accounting
I ask the AI to check whether the accounting records are complete. As a result, I receive a list of accounting records for which no document is available. This shortens the time I need for the advance VAT return by approx. 1 hour.
Training management
The AI can determine for me which training courses a customer has completed at the CCD Academy. This can also be done via the user interface, but is quicker in a request in German. The result can also be stored in an Excel spreadsheet, for example.
Service orientation
In the 90s, the idea of the Service-oriented architecture (SOA). It should be possible to integrate applications from different manufacturers in such a way that users can cover their business transactions across applications. The integration of order data in one application with accounting data in another would make it possible to automate processes across application boundaries. This idea hardly achieved any practical relevance as the integration was far too complex. Although there were standards, implementation was very time-consuming and only worthwhile for large companies. Small and medium-sized companies were practically unable to benefit from it.
Breakthrough of AI
Then something unexpected happens: suddenly AI is advancing in leaps and bounds, Anthropic defines the MCP standard, and things become possible that we dreamed of in the 90s. I can now talk to the AI about my accounting data and include receipts or bank statements. I can automate tasks without any programming effort. This suddenly makes integration possible even for clerks without in-depth IT knowledge. But first things first.
Model Context Protocol (MCP)
Anthropic, the manufacturer of Claude Code, had the idea of defining a standard protocol for integrating a wide range of applications. This is how the Model Context Protocol, MCP for short. Put simply, it works like this:
- An AI uses its MCP client and queries an MCP server for the list of tools it offers (tools/list).
- The MCP server responds with structured information, which essentially consists of text.
- The AI develops a plan on how it can use the tools offered by the MCP server to fulfill the required task.
- The AI instructs the MCP client to send calls to the MCP server using the necessary parameters (tools/call).
MCP is a protocol definition that enables us to make the functionality of an application available to any AI system. The only requirement is that the AI system must have an MCP client in order to be able to address the MCP server. And, of course, the applications to be used must provide an MCP server.
AI tools like OpenCode, Claude Code etc. naturally have an MCP client. And many developer tools such as IDEs, version control etc. have MCP servers. However, the integration of applications far removed from software development is an exciting field.
Current Clean Code trainings
We also run all seminars as closed company courses for you.
If you are interested or have any questions please contact us.
Practical examples
I use a desktop application for my bookkeeping: Monkey Office. This has an API that can be accessed via HTTP. The JSON format used is scary, but you can access almost all the data. In particular, you can access the posting records. A frequently asked question is: is there a document for every posting record? I could solve this problem using the programming interface provided by writing a program: simply retrieve the posting records and search for PDF files in a directory structure according to certain criteria.
But then, especially at the beginning, new requirements would constantly come around the corner, which I would have to implement individually in the program. For example, the monthly posting record for the payment of VAT does not have its own document. Or the clearing of a PayPal payment against the current account. There are many posting records for which exceptions would have to be defined in the program.
With the AI and an MCP server, the solution looks completely different. I describe the task to the AI in natural language. To do this, I write my requirements in a Markdown file. This is easy for me to write and easy for the AI to interpret. Markdown has established itself as the standard for various files used by AI tools. By storing my requirements in a file, I can transfer them to the AI again and again and ask it, for example, to carry out the document reconciliation for January 2026. The next month, I instruct it again and use the task description I created once. This can then be easily supplemented with further details that come to light over the months.
The key element here is the MCP server for accessing the accounting software.
Implement your own MCP server
I built the Monkey Office MCP server myself. To be more precise: I had it built in large parts by Claude Code and OpenCode. It wasn't the vibe coding of a beginner, but I had to tap the AI on the fingers in various places and give it tips. In the end, however, I had an MCP server that I can now use to access Monkey Office. It took me about 4 hours. The combination of Clean code and AI has led me to my goal.
As the AI can use the available tools, it independently determines that it can retrieve a list of postings. It also determines that it can filter, for example, according to a date range or specific debit or credit accounts. This means that I can easily formulate exceptions in natural language:
If the debit and credit accounts are from the 18xx range, it is a transfer posting and no document needs to be searched for.
I can write this in a few seconds and it works. If I had to implement this exception programmatically, it would also be an easy task, but it would take at least 10 - 15 minutes. And the best thing about it: such a natural language rule can also be written by someone from the specialist department who has no programming knowledge whatsoever.
However, despite all the euphoria, the result must of course be checked professionally. This is similar to programming. Here, too, I have to write tests to check the result. When using AI, my testing has so far been done „on foot“.
Potential
I see great potential here. By providing a standardized programming interface, AI tools can access practically any application. This will finally create the integration we dreamed of in the 1990s. No specialist knowledge is required. Quite the opposite: domain knowledge is paramount. The better I can describe my task to the AI, the better the result.
It remains to be seen whether MCP will remain the standard in the long term. It is currently the protocol of choice. Programming an MCP server is basically easy. I did this for Monkey Office using Microsoft's NuGet package ModelContextProtocol done. If an API already exists for the application to be connected, it's child's play.
Integrate MCP Server
Things become more difficult if you want to integrate the MCP server directly into your own application. This is only easy if the application has a clean structure. At best, all tools that are to be provided via MCP are already accessible via a method within the code base.
Here is another example. I have written a training administration system for the Academy in which we store all training courses. This offers our participants the advantage that we can provide a link where all relevant details about the training can be accessed. For example, the agenda, training times, exercises, sample solutions, the link to the Zoom meeting, etc. The application is implemented with ASP.NET Core and Blazor.
To create a new training course, I previously had to log in to the training administration and enter the relevant data in a form. This only takes a few minutes, but until now it could not be automated or integrated into other applications.
I have now implemented an MCP server in the training administration, which the AI can use to access the required data. So I can now formulate in natural language:
Create a Clean Code Developer Basics training course for customer xy. The trainer is SL and the times are 02.03./09.03./16.03. from 9 am to 5 pm.
It is a great relief that the AI finds out for itself that the three training dates must each be specified individually in the format YYYY-MM-DD HH:MM HH:MM. It creates the parameters for the MCP call independently and the training is created in just a few seconds.
Also fascinating: three IDs are required to create a training course: customer, seminar and trainer are specified via their respective IDs. The AI must therefore first find out which ID the named customer has. Furthermore, the IDs of the seminar and the trainer must be determined. The AI independently discovers that this can be done using tools from the MCP server and calls them up accordingly. Only when it has collected all the necessary data does it call up the tool to create a new training course.
The AI works out the solution to the task independently. This is where the great potential lies. It greatly increases productivity. I can suddenly complete tasks in a fraction of the time it would otherwise take. Above all, however, various applications can now be integrated. The next step could be the following:
Create an invoice for the training day 04.02.26. Submit the PDF to me for checking and ask me for approval. As soon as I have approved the invoice, email it to the customer.
Sounds utopian? Technically, this is no longer a challenge. As long as the required tools are accessible via MCP servers, the task can be solved by the AI. The important thing here is to use the so-called Plan Mode from Claude Code or OpenCode. This means that the AI first presents its plan of how it wants to implement the task and asks questions if necessary. This provides security. Ultimately, the details of such a task description should be stored in the system, e.g. as part of so-called Skills. Here, for example, there may be specifications for the email text or other important details. The more precisely the task is described, the better the solution.
Conclusion
MCP servers were perhaps originally planned to provide AI coding tools with the tools they need. But of course other use cases can also be realized with it. AI tools such as Claude Code or OpenCode run on my computer and therefore the file system is available to them. I can therefore include PDF files, for example, and thus carry out a comparison between accounting and a bank statement without writing a line of code.
Our task as developers is to provide the required MCP servers. In our training courses, we are happy to show you how to structure your application so that such tasks are easy to accomplish.
English 
German