What a new AI protocol means for journalists

There is a "rupture in journalism around AI" as media researcher David Caswell puts it. And this rift runs similarly through other sectors where people work with knowledge and words. Yet beyond the often heated debates about the usefulness and drawbacks of generative AI, which allow little room for nuance, there is bustling activity: Besides the large tech corporations, countless IT companies, startups, and individual enthusiasts are working on building an infrastructure for artificial intelligence (AI). Currently, artificial intelligence primarily refers to large language models (LLMs).

Through this activity, two areas have emerged in recent months that carry significant implications for journalistic work. Methods and approaches that were previously only possible with great effort, steep learning curves, or high costs are now becoming accessible: programming and operating complex software applications. Through LLMs, journalists' digital toolkits are now directly connected to an entire hardware store with its huge variety of equipment.

Coding agents

Using a coding agent means software is semy autonomously writing other software itself. This approach is also called "vibe coding" — presumably because it involves a rather fluid, iterative, and dialogic way of working. The user engages with the oscillations of large language models — with LLMs, there's always an element of chance involved.

These tools enable something that previously posed difficulties for people who couldn't program themselves: Implementing their ideas in practice. Yes, you still need a basic understanding of digital technologies and software coding. But until recently, digital projects, whether designing a new website or creating small tools, usually required a human software developer. For example, to collect information via crowdsourcing or to automate the processing of recurring datasets. And for more complex projects, designers were also needed for the interface (User Interface, UI) and functionality (User Experience, UX).

Coding agents now take on this work: From developing the UI, the structure, setting up a database, to publication (deployment). The resulting code "belongs" to you; everything is based on open-source software. This means prototypes and ideas can be further developed elsewhere. The tools are particularly suitable for web applications (web apps).

For instance I have used coding agents to retrieve real-time data from German railways via their official OpenData interface (API), to build a tool for collecting, visualizing, and analyzing delay data. Or to quickly craft a streamlined interface that allowed me to easily read and search through an extensive archive of messenger data. Finally, from the idea of obtaining a simple tool to quickly create transcripts of long radio pieces, we built a tool that was completely developed by coding agents without human developers: DIVER summarizes podcasts and newsletters, analyzes them, and provides users with overview reports and recommendations for podcast episodes and newsletter issues.

It should be clear: Data-sensitive applications intended for use in critical areas should still not be developed without the involvement of professional human programmers.

However, it is also clear that the increase in performance of these tools will likely continue for some time. This means that even larger software projects will soon be implementable quickly and cost-effectively.

Meanwhile, there are about two dozen providers whose products differ in nuances, approaches, and focus areas — a selection can be found at the end of the article. Almost all offer free entry or even a daily free budget that can be used. Typical prices for the first payment tier range from 15 to 20 euros per month. Google recently entered the ring with its "Firebase Studio;" LLM manufacturer Anthropic agreed to work with Apple on a coding agent for the Xcode programming language. And OpenAI recently purchased the development environment "Windsurf" for $3 billion, which can also be used as a coding agent.

Model Context Protocol

In spring 2025, a crucial component of the AI infrastructure took shape: The MCP protocol. The open "Model Context Protocol" was introduced by Anthropic, known for its Claude chat, in late 2024. OpenAI, Microsoft, and others have since joined. So what does it do? "Think of MCP like a USB-C port for AI applications," says the opening line on the standard specification website.

Or to illustrate it differently: Those who have seen the film The Matrix might remember this scene: Trinity and Neo stand within the digital world of the Matrix in front of a helicopter. Neo asks her: Can you fly this? She answers: Not yet. A few seconds later, she has downloaded the corresponding skill and can pilot the helicopter. MCP reads the handbook to steer complicated software for you.

MCP positions itself as an intermediate layer between existing software applications and data sources on one hand, and large language models on the other. This means programs like Photoshop, 3D software like Blender and Cinema 4D, and music programs like Ableton, which come with complex user interfaces and steep learning curves, can now be operated via language: Users describe via chat, spoken or written, what they need or want to change. The MCP then controls the user interface of the software. It has essentially read the manual and knows how to operate it. The result of its activity can then be observed immediately and, if necessary, tested. In this respect, the MCP enables something comparable to the WYSIWYG approach​​​​ (what you see is what you get) in graphic programs, etc.

This should be particularly interesting for data journalists. There is already an MCP solution for RStudio, an application for using "R" for statistical analysis. It can be controlled via prompts like "Load the dataset and create a scatterplot of X vs. Y with a trend line." It is only a matter of time before visualization tools like DataWrapper can be controlled via MCP via chat to create map and data visualizations.

What does this mean for journalists?

As often happens, such technological change presents a double-edged sword. On one hand, it empowers individuals and resource-poor newsrooms directly with working methods that were previously unaffordable: Extensive research, persistent monitoring, complex analyses, and sophisticated digital formats. Used skillfully and knowledgeably, this can significantly increase the quality of journalism.

On the other hand, it's also clear that these new tools will be used to obscure and complicate research: For example, to create credible forgeries of digital content. And the major AI corporations see the future in digital employees and digital twins — digital representatives of a person. It's still unclear exactly how this will take shape. Amazon, for example, recently showed how an LLM agent can operate directly on webpages in a browser. And Google introduced the open A2A protocol, which is intended to regulate the communication and coordination of AI agents with each other, regardless of who manufactured them.

What kind of relationship should journalism maintain with digital representatives?

How are their "actions" to be evaluated? How can responsibility be clarified and assigned when concrete effects on people occur through the digital activity of a digital twin? These are all open questions. The new possibilities offered by programming agents and the simplification of operating complex software applications will help find answers to these.

List of coding agents

Magic Patterns (User Interface)
V0 (for beginners, but still powerful)
Lovable (beginners and advanced users)
Replit Agent (beginners and advanced users)
Windsurf (professional users)

Lorenz Matzat is a journalist and software producer working on LLM tools for journalism and knowledge retention. In 2017, he co-founded the NGO AlgorithmWatch, where he led its Research & Development efforts until 2022.