This piece was originally commissioned by and published in Dutch at Villamedia, June 24, 2026
Political editor Hans keeps hearing that something has really changed in the field of AI… for the umpteenth time. New software such as Claude Cowork, OpenAI’s Codex, and MCP servers are said to be true game changers because they connect digital services, program functioning apps on their own, and automate tedious workflows. Hans: “I already struggle to get ‘ordinary’ chatbots to do what I want. Help me, Laurens: do I really have to delve into this now?”
During one of many journalism panels on AI that I attended this year —this time at the International Journalism Festival in Perugia— moderator David Caswell and his panelists used a particular acronym nine times. That three-letter-acronym, MCP, seemed to be entirely new to many people in the room—and probably to you too! As a teacher, my heart always bleeds a little when a speaker assumes certain knowledge as basic, but loses the attention of a group by not reading the room properly.
After the reader: what comes next for news in an AI-first world? Panel session with David Casswell @ International Journalism Festival Perugia
In short, what *is* MCP?
The letters stand for Model Context Protocol. Without getting too technical, in essence, it comes down to the fact that, based on an agreed set of standardized rules, existing websites and apps can be summoned through AI chatbots. The new protocol enables users to query data and content from those apps (such as spreadsheets, documents, and images – and what have you) by using natural language. You can then use your chatbot of choice to act on those answers, exploring it further with follow-up questions, by generating visualizations or building new apps on top of the chatbot’s output.
For journalistic applications, you could consider: ‘retrieve all mentioned geo-locations from your paper’s innovation section-archive, and position them on a geographical map of Europe.’ Or you could ask your chatbot to do the following via Good Tape, a Danish AI transcription service widely used by journalists: ‘list all philosophical quotes in my five most recent interviews. Pay attention to references to Hannah Arendt, Kant, Popper, and Aristotle.’
Another example: when analysis tools embrace MCP, you will soon be typing, ‘Hey chat, which user need increased our reader-conversion most, over the past nine months?’ Silja, Product Marketer at Good Tape, says that their MCP (now in public beta) has changed the way she works with transcripts. “I asked Claude to turn user interviews into concrete insights. That one assignment easily saved me an afternoon.”
Panel moderator Caswell predicted that every journalist will soon be running their own em-cee-pee server. He encouraged the audience to try to free up some on a Saturday evening, pay 18 euros to Anthropic and ask Claude to assist setting up the MCP-connection of choice.
Why would you run an MCP server? Because it is a logical next step in the evolution of AI, Caswell and other proponents believe. After all, we have been ’talking’ to our chatbots for almost four years now. With this protocol, it is time our sites, apps, and content management systems start doing that as well. The next step is for machines and apps to simply communicate with each other and perform tasks mutually, on their own. So humans can step aside. Let the machines do the dirty work, at least—that’s the promise.
Admittedly, it sounds both frightening and promising.
Caswell’s appeal convinced me. Upon returning home, I slavishly followed his call and purchased a paid Claude subscription. I read up on MCP and was ready for my experiment: making machines talk to each other and work together … under my eye!
Experiment in Lycra
To succeed, I needed three things: a goal, data, and domain knowledge. These turned out to be within reach. Tomorrow (July 5), along with thousands of other MAMILs (Middle-Aged Men In Lycra), I will cycle the Maratona: 138 kilometres up- and downhill in the Italian Dolomites. Not a very attractive idea to many; same for untrained amateur cyclists.
I saw this adventure as a perfect opportunity to combine my hobby with new technology and my journalism work. A handy bonus is that Strava, the well-known running and cycling app that I already use, recently fired-up their own MCP server.
Caswell was right: that very same evening I started the experiment, I managed to query my Strava data in a casual, conversational way using chatbot Claude. I easily generated a small table showing how my climbing fitness was improving. The bot had no great motivational powers, though… After I mentioned the dry tone, Claude’s adjusted response didn’t improve. It made blunt statements that you wouldn’t accept from a meat space trainer. I was repeatedly told that I was not losing weight fast enough; after all, the fewer kilograms I carry, the easier it will be to get to the top.
What my synthetic sports coach also kept urging me to do is a training session with at least two thousand meters of elevation gain. By comparison: Alpe d’Huez has about eleven hundred. While I live in the Dutch Veluwe-region, which is rather hilly given our mostly flat country, travelling to mountainous territory was not an option at that point. “Have you seen my calendar, dear?” I replied. – “If you connect those, I will include that in the advice.” I considered that too dangerous from a privacy and security perspective.
Mixed results
That I managed to produce a sensible overview in natural language – without perusing boring spreadsheets, broken formulas, or complex database queries—is truly fantastic. The possibilities for new applications quickly opened up in my brain. I could now understand why the AI-pilled people are excited. I know the training data my bot uses very well, and that it’s accurate because I’ve just generated it myself, half an hour ago, outside in a damp forest, in the real world, by making a physical effort.
Not only that, but I am genuinely surprised when the chatbot manually recognizes previous indoor simulation rides with names like Campolongo and Giau from my Strava-history. Something I didn’t explicitly tell Claude, but relevant nonetheless: these are the names of mountain passes I’ll have to climb during the Maratona ride.
And, of course, Claude being a chatbot, it upholds its questionable reputation when it comes to language proficiency. After four training sessions in one week, it concludes that I “really jailed it” this week. Say what?!
Is this Big Tech’s latest Trojan horse?
So much for the pleasant surprise and practical benefits of these new AI capabilities. What is possible for cycling can also be applied for use of news content. When newsrooms start providing their own MCP-servers, users can call up a financial or sports updates every day, tailored to their particular interest and knowledge level, filtered for markets and sports that are important to them. And then happily chat away about their favourite’s Men’s World Cup Soccer goalkeepers.
Remember Hans, who asked if he should get involved in these kinds of experiments? Well, some of it seems to have some potential when it comes to convenience for our audiences. But it does also mean that our stories become even more ’trapped’ in chatbot environments. Who subsequently increasingly capture our audience and their attention, and are in good position to capitalize on that.
Seen in this light, a news org offering an MCP is more akin to rolling out a red carpet for Big Tech’s next Trojan horse. Besides this consolidation of power concentrations, there are two other dangers that are economic and editorial by nature. That’s because by now, your stories live out there on someone else’s, not-well governed digital platform and have been scraped without consent, credit or compensation (triple C’s), as raw material for large language models. As AI-companies start to happily connect their software and platforms with our apps and services, our content will flow through many more external touchpoints, again mostly without the triple C’s being addressed. Unless publishers close deals with every new AI-provider, journalists once again do not benefit financially
from this new technology. (An other thing I learned in Perugia, is that it’s worth more than a pretty penny to most publishers if they would only know how and where parts of their content lives within these chatbot-surfaces, aka as part of generated responses to users)
But more importantly, there’s an editorial danger regarding MCP: we – or our news publisher – are being kept from (or negotiated out of) a position to manage the additional data streams and processing points ourselves. In doing so, we create new vulnerabilities to the integrity, security and privacy of our editorial data, our sources, and our stories.
On its surface, it seems convenient if Good Tape connects your transcripts to your chat account via MCP, but who says OpenAI or Anthropic (or any other 3rd party) will treat the contents of that interview carefully and non-commercially? What if your chatbot or someone else’s MCP server gets hacked? What if access to a model or platform is cut off by a presidential Executive Order?
Who gets the final call?
On a technical and functional level, MCP as a protocol is pretty interesting. New opportunities and idideas,reated through connecting services, reveal themselves immediately at the first prompt, and building something usable and useful is just a natural language sentence away. Also, the fact that the industry somehow managed to arrive at a shared protocol at all, is a powerful signal that universal agreements, reached through cooperation are still possible in this day and age.
But, as always, technology works primarily to the advantage of the already powerful.
AI fanboys— and people with an agenda, or selectively informed individuals—tend to draw conclusions too quickly and superficially based on their non-representative, personal experiences. They do not take the socio-economic context of every technology (it always has to ‘live’ somewhere, right?) into account in their promises-and-prediction-laden posts on LinkedIn. They view the output of large language models as unequivocally valuable; always accompanied by the obligatory disclaimer that some human still needs to check somewhere, mostly at the end of the loop, to gauge to what extent the slop is sound. Just like the language models themselves, most of their promises eventually hold no water upon scrutiny.
But hey, I am no better myself. AI evangelists demonstrate their confirmation bias by often providing somewhat contrived examples and by pointing to simplistic benefits that primarily align with their existing beliefs. I do the same thing, but in the opposite direction. If I detect some shortcoming of an AI application, I wave away actual utility or, easy as that, an entire category. Always accompanied by my predictable spiel on the saving of humanity, moral uplift, and democracy-preserving arguments.
However. Many tech bros make extraordinary claims without providing extraordinary evidence. Consider, for example, the predicted mass unemployment caused by AI, or the ever-imminent AGI superintelligence. They advocate that using AI tools in a smart and responsible way, will make the difference between the haves and the have-nots.
Ask questions, perform a speech act.
Carissa Véliz, philosopher at the University of Oxford, has written in her latest book ‘Prophecy’, that a prediction can be descriptive or prescriptive. A weather forecast does not influence the weather. The promises of Big Tech, on the other hand, are ‘veiled commands’. If we believe them and act accordingly, they create self-fulfilling prophecies.
Fortunately, language belongs to everyone, and journalists make prescriptive statements just as much. However, we use the language we employ to reinforce our own futures too infrequently. In this context, Véliz refers to the concept of the ‘speech act’ (as originated by philosophers J.L. Austin and John Searle): an action that takes place because something is said with a certain goal in mind, e.g. ‘clean up your room’. Perhaps that is what MCP servers are for: they help us to perform technological actions through language.
To get back to Hans’ question: first ask yourself whether and how new AI software serves editorial purposes. If you start there, you won’t need to try out most AI programs yourself. Use language to set people and ideas in motion, not machines.
Tomorrow in the Dolomites, when I climb that final mountain, legs burning, drenched in sweat, exhausted, I might summon God or the devil… but I really need to this on my own. You’d wish Claude could experience something like that.
Note regarding this experiment:
Against my better judgment, I developed some FOMO regarding the latest developments in AI. Precisely to avoid simple counterclaims or criticism from the comfort of my armchair, I allowed myself to be tempted to join the camp of the chatters, vibecoders and agentics.
I do (and did) have some moral difficulty undertaking this experiment. For quite some time now, I have used virtually no generative AI software, no large language models nor chatbots, due to their extensively researched and demonstrated problematic aspects. I view this experiment as an exception that -hopefully— serves some broader benefit.
–
This piece was originally commissioned by and published in Dutch at Villamedia, June 24, 2026