Opening Note: This is the first edition of a report on the use of the Model Context Protocol (MCP) in journalism, produced as part of the Jor-MCP project. Ambiental Media develops Jor-MCP, an open-source MCP server designed for journalism, and therefore has a direct interest in the topic. An expanded version, featuring a comparative analysis of the case studies, will be published shortly.
What Is MCP
Written by Giulia Afiune
Updated on October 30, 2026
The Model Context Protocol (MCP) is a communication standard between systems designed to optimize connections between artificial intelligence tools and the outside world. With MCP, AI agents can access, for example, third-party databases and apps from other companies.
Developed in 2024 by Anthropic, the creator of Claude, the MCP protocol was quickly adopted by other tech industry giants.
Since its launch, the MCP has become a standard language of sorts, spoken and understood by various tools within the AI ecosystem. The MCP enables AI agents to “talk” to other sources to find relevant information and communicate with one another to perform complex tasks.
This adaptability has earned the MCP protocol the nickname the USB-C of AI. In the past, each cell phone brand used its own type of charger that didn’t fit devices from other manufacturers. Over time, the USB-C cable became popular, and in 2022, it was established as a mandatory standard in the European Union.
Just like a USB-C charger can be used with various cell phone models and brands, when an AI model uses the MCP protocol, it can connect to various tools and data sources, provided they also follow the protocol’s rules.
What Does the Acronym MCP Stand For?
In the field of Artificial Intelligence, the term “context” refers to the set of background information that AI tools use to generate responses for the user. Context includes the prompt entered by the user, their previous interactions with the tool, and other external data sources that help the system understand the situation and the user’s request. It should not be confused with the data used to train the model.
Therefore, the Model Context Protocol is simply a protocol that standardizes how AI models receive context so they can operate more effectively and efficiently.
How the MCP works
The MCP is an open standard: its specification is public and can be implemented by anyone, and the reference libraries are open source.
The main components of the MCP architecture are the host and the server. The host is the AI tool that receives user requests and seeks to respond them — it could be Claude, ChatGPT, or Gemini, for example. The server is the program or external database that provides context to the AI tool so it can respond to user requests.
Companies need to make their databases available on MCP servers so that AI tools can access them via the MCP protocol.
Client applications
Chat interface
Claude Desktop, LibreChat
IDEs and code editors
Claude Code, Cursor
Other AI applications
Systems and superinterfaces
MCP
Standardized
protocol
Data and file systems
PostgreSQL, Google Drive, etc.
Dev tools
Git, Sentry, etc.
Productivity tools
Slack, Google Maps, etc.
External services and tools
Client applications
External services and tools
What the MCP Is Not
The MCP is not a large language model (LLM), such as ChatGPT or Gemini, but rather a set of rules that these applications use to communicate with the outside world. MCP is also not a tool that operates without human supervision, although its use facilitates the development of autonomous AI agents. Finally, although widely used in the AI world, the MCP protocol does not replace other data-sharing formats, such as REST APIs, NewsML-G2, and RSS feeds.
Advantages of the MCP Protocol
Before MCP, LLMs and AI agents operated based on the knowledge passed on to them during model training. These tools also performed internet searches and data scraping, but this returned a large number of irrelevant results that required significant processing power and high costs.
Furthermore, LLMs such as ChatGPT and Gemini had difficulty accessing up-to-date information in real time, such as daily news, and depended on custom integrations to work with private information, such as corporate databases and internal company systems.
The adoption of MCP as an industry standard has enabled AI agents to access specialized knowledge bases in a standardized manner, without the need to create a specific connection to that source. This means that AI agents can now think and perform tasks based on this external knowledge, which expands their ability to carry out more complex and personalized tasks for individuals and businesses. In practice, the widespread use of the MCP protocol has accelerated the development of smarter and more efficient AI tools, paving the way for the creation of agents that perform actions autonomously.
How the MCP Protocol Is Used
The standard established by the MCP has made it possible to develop new AI agents to automate repetitive tasks and optimize workflows in professional environments.
One example of individual use is that you can configure your ChatGPT to connect to your Google Calendar’s MCP server, since both use the MCP protocol. This means you can directly ask ChatGPT to create and delete events in Google Calendar.
Within a company, Microsoft Copilot can be used to access multiple internal files via the MCP protocol, allowing users to analyze those databases using natural language.
Why MCP Matters for Journalism
MCP is a key component of the infrastructure needed to adapt newsrooms and journalistic content to the AI era. However, the adoption of the MCP protocol in the journalism industry is still in its infancy. Jor-MCP was created precisely to accelerate this transformation.
Media outlets around the world have already started using tools based on the MCP protocol in a number of different ways: to assist journalists in the reporting process, to search for stories in their archives, to locate relevant images for video editing, and to make news content available for consumption by their readers' AI agents.
In Chapter 2 of this report, we present use cases for MCP in a news organization, Reuters News Agency, and at a scientific publisher, Wiley. Based on interviews with representatives from each company, we address topics such as development and infrastructure, practical day-to-day applications, monetization models, editorial integrity, and next steps in the development of these tools.
Security
The rapid spread of MCP in the AI industry has also raised concerns regarding the protocol’s security.
The document cites a number of vulnerabilities in the design of the MCP protocol. One of them is that user identification and authentication to access MCP servers is not mandatory, which can facilitate unauthorized access and data leaks.
Furthermore, since the protocol allows the transmission of text snippets without validation, malicious code can be embedded within user requests (a practice known as prompt injection). The NSA also warns against inefficient approval workflows, as changes to the capabilities or permissions of an already connected MCP server can occur in the background without requiring new user consent.
Another risk is exposure to prompt storms that overload the server. The lack of detailed logs of command execution also makes it more difficult to trace incidents and identify anomalies.
For these reasons, the NSA recommends that developers go beyond the security measures suggested in the protocol and take additional precautions when creating MCP-based tools and systems.
The recommendations include limiting access and restricting permissions, separating data storage and processing environments, segregating and protecting sensitive data, using only regularly updated MCP tools, validating the parameters of both user-provided messages and those returned by the system, and implementing a logging system for executed commands.
Anthropic launched the MCP on November 25, 2024, to address a connectivity issue with AI models.
Prior to that, despite rapid technological advancements, AI models—though powerful—were still limited because they struggled to connect with users’ real-world information sources. “I have like [an] amazing AI system, this brain, but it's kind of put into a jar and can't reach out to the world. And so, it was very clear, very early on for me that we need some form of connectivity,” explained David Soria Parra, a member of Anthropic’s technical team and one of the co-creators of the MCP, in an interview for the Stack Overflow blog.
Before the MCP, whenever it was necessary to connect AI models to other tools and data sources, developers had to manually create a new, specific integration and ended up copying and pasting code and prompts between applications. According to Parra, it was the frustration he and his colleagues felt with this repetitive and error-prone process that motivated the creation of the MCP protocol. “What we are effectively trying is building a plug-and-play system like the internet has very rarely seen before, where you can take any client and any server and connect them freely with each other,” adds Parra.
This proposal was so promising and strategic that, in about six months, four AI giants (Microsoft, OpenAI, Google, and AWS) adopted the protocol created by a competitor, effectively making MCP the industry standard in the tech sector.
Initially, MCP could only be used with Claude, Anthropic’s AI application. Microsoft was the first competitor to launch a version of its AI application, Copilot Studio, with MCP integrated, on March 19, less than four months after the protocol’s launch.
The following week, on March 26, 2025, OpenAI announced that ChatGPT and other company products would begin to natively support MCP. “People love MCP and we are excited to add support across our products,” said CEO Sam Altman in a post on X.
The race to adopt MCP culminated in an unlikely institutional decision. On December 9, 2025, Anthropic donated the MCP protocol to the newly created Agentic AI Foundation (AAIF), co-founded alongside OpenAI and the fintech firm Block with support from Google, Microsoft, AWS, Cloudflare, and Bloomberg — all of which are represented on the foundation’s board.
As a result, governance of the MCP and other similar open-source projects is now shared among the industry giants under the neutral umbrella of the Linux Foundation, ensuring transparent and collaborative evolution of the protocol and preventing its future from being dictated by a single company.
“Bringing these and future projects under the AAIF will foster innovation across the agentic AI ecosystem and ensure these foundational technologies remain neutral, open, and community-driven,” Anthropic explained when announcing the donation.
Timeline:
November 25, 2024
Anthropic launches the Model Context Protocol (MCP)
March 19, 2025
Microsoft launches the first version of Copilot Studio with integrated MCP
March 26, 2025
OpenAI announces that ChatGPT and other company products would adopt MCP
April 9, 2025
Google announces that Gemini and other company AI models would adopt MCP
May 29, 2025
AWS launches its first MCP servers to assist developers
November 2025
Ambiental Media begins developing Jor-MCP with funding from Codesinfo – the Journalism Innovation Fund, an initiative carried out by Projor with support from the Google News Initiative
December 9, 2025
Anthropic donates governance of the MCP to the newly created Agentic AI Foundation
February 2026
Beginning of the second phase of Jor-MCP’s development as part of the JournalismAI Innovation Challenge, an initiative by Polis, the journalism think tank at the London School of Economics and Political Science (LSE), with support from the Google News Initiative
May 31, 2026
Ambiental Media launches the beta version of Jor-MCP at the 3i Festival in Rio de Janeiro
October 30, 2026
Official launch of Jor-MCP
How Is MCP Used in the Real World?
Written by Giulia Afiune
Updated on October 30, 2026
Case 1: Reuters
Founded in 1851, the Reuters News Agency is the world’s largest news agency. Reuters licenses news, reports, stories, and multimedia journalistic content to media outlets around the globe in 12 languages.
On July 8, 2026, the company launched the Reuters MCP server, a new agency-controlled way to distribute its content to subscriber media outlets, as Alphonse Hardel, head of the Reuters News Agency, explains in an exclusive interview with Ambiental. “The biggest opportunity is that AI agents become a new channel for our journalism, and one we can govern,” says Hardel.
The Reuters MCP server makes the agency’s content (text, photos, videos, graphics, and audio) available to be searched, read, downloaded, and edited by AI agents, search tools, editing platforms, and internal newsroom systems that were already in use by its subscribers.
“A lot of research and production now happens inside AI tools, so we wanted to remove the friction of using multiple platforms and allow our content to be in the AI workflows our customers are already using”, explained the head of the Reuters News Agency.
One of the goals of the Reuters MCP server is to improve the efficiency of newsroom workflows, freeing up time spent on manual tasks so that journalists can focus on more complex investigations. “For us, that puts Reuters inside more of our customers’ workflows. For them, it means editors spend less time on routine searching and more time on stories. Early uses such as fact-checking, daily briefings and image matching show how far that can go,” says Hardel.
Target Audience
The Reuters MCP server is part of the service offered to subscribers of the Reuters News Agency, so the primary target audience is journalists at client newsrooms. A version of the tool is also being implemented inside the Reuters newsroom for its own journalists.
Use Cases and Features
Hardel explains that the Reuters MCP is used in similar ways by partner media outlets and within the Reuters newsroom. “AI agents can automatically surface and insert relevant Reuters visual assets directly into drafts as journalists write, eliminating manual media searching and freeing up journalists’ time to spend on reporting and newsgathering. AI agents can efficiently repackage stories into various formats. And AI agents can surface clips to quickly create a video edit,” says the director.
The server itself allows AI agents to perform searches (finding content and filtering it by media type, date, geography, and language), retrieve items (the server returns the full text and metadata of a news story), and download (delivers content in the format the user needs). “We prioritized these because they match how a researcher or producer actually works [with our content]: find it, check it, take it,” explains Hardel.
The Reuters MCP is also being optimized to work with multimedia content. The server provides live streams so that client media outlets can discover and activate real-time coverage from Reuters, which is especially useful for broadcasters. In addition, the server is integrated with ShortCut, a video editing platform from CuttingRoom, which allows journalists to search for and edit videos from the Reuters News Agency archive using natural language directly in their browser.
Results
The clearest gain of the Reuters MCP so far, according to the news agency’s head, is the reduction in “integration effort.” “Customers don’t need to ingest and maintain a full copy of our content, and they don’t need to build rights logic or a separate pipeline to match entitlements,” says Hardel. This results in improved workflows because “a lot of manual tasks, such as monitoring the Reuters output can now be automated and ready for presentation to an editor at any time.”
Another advantage reported by users is time saving, which are crucial in journalism. “We’ve heard from customers that their turnaround time is faster, allowing their journalists to focus on editorial decision rather than finding relevant content.”
The director explains that the metrics monitored by Reuters to measure the tool’s efficiency are: adoption (how many customer accounts are actively using the MCP, as well as the number of connections and searches); quality (whether the user’s first search yields a useful result or a download); and time saved for each pilot user (compared to the average baseline time agreed upon in advance).
Hardel notes that it is not possible to say whether using the Reuters MCP server has helped reduce newsrooms’ AI costs, since clients themselves choose their AI tools and determine how data is stored and processed.
Business Model and Monetization
The Reuters MCP server is a tool integrated into the system provided by Reuters News Agency to its clients, designed to add value to existing subscriptions. “Licensing remains the foundation of how Reuters News Agency monetize our content, and agents are another way our customers receive it,” says Hardel.
He notes that Reuters is still exploring ways to offer its content directly to readers through AI agents. However, access to the MCP server is not currently sold separately.
Hallucinations and Editorial Integrity
Maintaining editorial integrity is a key concern when processing journalistic content with Artificial Intelligence. In the Reuters MCP, this risk is minimized because the server does not generate news stories. “It retrieves what our reporters and editors have already reported or verified under the Trust Principles, so the journalism itself doesn’t change. Because the server returns full text along with the metadata, the agent is working from the story as published, not a summary of it,” says Hardel.
In addition, he explains that each item carries a unique identifier code, allowing any content that comes out of the tool to be traced back to the original story and properly credited.
The codes that identify the media type and subject also prevent the model from inventing new categories. “At present, the server returns the content and standard metadata, with no added story background. Human editors still decide what gets published”, Hardel assures.
Infrastructure
Since Reuters content is licensed to other outlets, it was already categorized and identified. As a result, implementing the MCP server did not require major adjustments. “We didn’t start from scratch. Our content already carries unique identifiers, detailed metadata and standard media and topic codes, and the server relies on all of them,” says Hardel. However, there was a bottleneck with video content. “Shot-level and concept-level search isn’t yet in our primary search engine. Closing that gap is the next piece of work.”
Hardel also clarifies that the MCP server is simply another way to distribute Reuters content and does not replace traditional formats such as APIs and RSS feeds. “A customer with a system built around a feed or API will keep it, and for bulk ingestion or a system of record, the traditional route is still the right tool” he says. “I wouldn’t call any traditional format obsolete. What’s changing is where people start, and increasingly that’s inside an AI tool.”
There are situations, however, where using the MCP server is more appropriate. “MCP excels when someone needs something specific and needs it now: find the right clip, check a fact, pull a story into a draft,” he explains.
Security Measures
User access permissions were Reuters’ primary security concern when implementing the MCP server. “Our content is licensed, and an agent should never see material a customer hasn’t paid for. Entitlements are applied automatically on every request, so customers can’t reach beyond their agreement,” says Hardel. He adds that the news outlet’s websites are protected against unauthorized scrapers.
Next Steps
Alphonse Hardel, head of the Reuters News Agency, reveals that Reuters’ MCP server will soon feature a new semantic search capability for text, images, and videos — which will also be searchable in greater detail by shots and concepts.
In addition, Reuters is developing connectors for leading AI apps, such as Claude and ChatGPT.
The agency is also exploring the possibility of delivering content directly to readers via AI agents.
Founded in 1807 in New York (U.S.), Wiley is now a global publisher responsible for publishing thousands of scientific journals, books, and academic publications across all fields of knowledge. In addition to making its scientific collection available in an online library, Wiley also operates as a technology company offering solutions to other publishers and the academic community at large.
In October 2025, Wiley launched Scholar Gateway, an AI assistant for scientific research that performs searches on an MCP server containing Wiley’s collection of 8 million peer-reviewed articles, according to the platform's website.
According to a study conducted by Wiley itself, 62% of researchers were already using AI for tasks related to research and academic publishing. “We were seeing increasing numbers of researchers using ChatGPT or Claude or other tools, starting to do quite a lot of work in those, but not connecting to the scholarly materials that their library subscribed to. That's fundamentally a problem and that's what we were looking to solve,” explains Chris Reid, senior director at Wiley.
While assistants like ChatGPT and Claude, which are not connected to specialized sources, search for information on the open internet, Scholar Gateway restricts its search to Wiley’s knowledge base of publications, which, according to the company, results in evidence-based answers.
“It’s been a real success,” says Reid. “We've interviewed a lot of users to say this is exactly the problem they were looking to solve. So that's been very, very rewarding.”
Target Audience
The tool’s target audience consists of researchers and scholars who use AI for scientific research. Scholar Gateway is available as a connector in Claude and Mistral Vibe and as a plugin in ChatGPT.
Use Cases and Features
The primary use of Scholar Gateway is the discovery of scientific content, as the tool allows semantic search across more than 8 million peer-reviewed articles published in the Wiley collection. Each search returns a result that includes the author’s name, the date and journal of publication, and a DOI, a unique identifier for scientific publications.
According to Chris Reid, while 85% of users employ Scholar Gateway for discovery, 15% are advanced users who utilize the tool to create multi-agent systems or operate autonomous labs. In these cases, the researcher formulates a hypothesis and configures an agent connected to the MCP to run multiple experiments overnight and analyze whether the data supports it. “There's lots of 'ifs' and 'buts' around this, but ideally, it [could be] an accelerant to scientific discovery. We are seeing that in quite a small but growing number of use cases,” says Reid.
Results
Initially launched as a pilot project, Scholar Gateway was made official in September 2026, in part due to positive user feedback. The company informs that more than 90,000 researchers have already registered on the platform, which reached the 1 million query milestone in mid-October 2026, according to Chris Reid. The Wiley director reported that 60% of searches were related to Science, Technology, Engineering, and Mathematics, while 40% were in the Social Sciences.
The key metrics monitored by Wiley are the number of users (counted per authentication), usage volume, and engagement with the content (how the user utilizes the results). Another metric observed is how the AI agent searches the Wiley database, that is, how many and which questions the LLM asks the MCP server. The original question asked by the user to the LLM is not visible to the Wiley tool.
Reid emphasizes that measuring usage volume through the MCP server can be highly relevant for news organizations. “One of the big changes in the media landscape is that most media publications have lost track of usage, either because it's massively inflated by machines or robots, or because if someone's on Google overviews, they read an overview and never click through to your article, so you cannot record what's happening there. For us, through Scholar Gateway we can record usage, and that's actually a really important measure.”
According to Chris Reid, some advanced users have reported that using Scholar Gateway has been transformative, as the tool helps them significantly reduce the time it takes to produce an article or research paper. Another piece of feedback received was the reduction in token usage for a search, since there’s no need to spend them processing PDF files — a format that’s highly inefficient for artificial intelligence.
“If you're an enterprise, you probably have Claude and you're paying for usage, so you want to be more token efficient. Sending a machine into an environment built for humans isn't very token efficient,” argues Reid. The senior director at Wiley believes that the pressure for token efficiency will grow as more companies adopt Scholar Gateway.
Business Model and Monetization
Scholar Gateway can be used on a free plan, which requires only registration and is limited to 30 searches per month. Launched on September 28, 2026, Scholar Gateway’s paid plan is designed for frequent users and offers up to 500 searches for US$15 per month.
Hallucinations and Editorial Integrity
Wiley has no control over what Claude, ChatGPT, and agents from other companies do, clarifies Chris Reid, “but we can certainly improve” the results. He explains that Wiley takes steps to segment and label its content to deliver the most accurate results possible and also suggests that the agent include proper attribution in its responses, which does in fact happen in most cases.
Infrastructure
The MCP is the backbone of Scholar Gateway and was the approach chosen by Wiley because it is a simple, straightforward, and widely adopted protocol. However, Reid explains that Scholar Gateway is MCP-agnostic. “We're really focusing on the problem. If MCP happens to be that solution, great. If Web MCP is the second stage, then fine. If a better standard emerges, we’ll adapt to it.”
Reid explains that, although academic writing has followed the same standard for a long time, Wiley had to adapt the way its content was structured so that AI could interpret it correctly.
“We optimize our content. We have a vector database and a strategy where we chunk each article into a significant number of chunks. We apply metadata against each of those, we vectorize that, and then we have a semantic search tool in our MCP, which then queries it. And then we optimize that tool as well to make sure that, if a user is in ChatGPT, they're getting a good result,” explains Reid. The director explains that the “secret sauce” Wiley uses to segment articles cannot be revealed, but he assures us that a key part of the method involves using mathematical models to establish relationships between different sections of the article.
This mechanism allows Wiley to better understand how users — both bots and humans — are utilizing its content. One interesting finding was the importance of the methodology sections in scientific papers.
“Scholars don't like writing them, they're often not brilliantly done, or are sometimes hidden away in supplementary data, but actually are really useful for a machine, particularly if you're looking at lots of methods, trying to create and recreate an experiment or reproducibility, all these sort of things,” adds Reid.
The senior director at Wiley notes that, with LLMs and chat interfaces, there are countless use cases, and MCP won’t always be the best choice for all of them. “Lots of use cases are still emerging, and lots of people are doing [research] in different ways. We're very much in the interim period and we're probably gonna be in an interim period for a long time [before we] settle down into a place: we use this for this, and that for that. Until then we have to be fundamentally as flexible as possible."
Wiley continues to offer solutions based on other technologies for specific use cases, such as an API for text and data mining that provides large amounts of data in PDF format for non-commercial use. According to Reid, the challenge is deciding where to invest the company’s resources. “How do we continue to support standards which are still really important for some audiences, while understanding more of our investment has to go into different areas? That's a fairly tough question,” he reflects.
Security Measures
One security concern during the Scholar Gateway development was preventing the MCP server from being used to perform a massive content harvest to clone Wiley’s collection, says Chris Reid. To mitigate this risk, he explains that a series of rules and request limitations (rate limiting) were implemented.
Another security measure was to allow the Claude connector to only read and not edit content, to prevent attacks such as prompt injection.
Wiley decided to block its website to AI models' training bots, thereby avoiding massive server costs. “We block the training bot, but we allow the user bot in because that's delivering on user queries. We wanna make sure our content is discoverable,” adds Reid.
Next Steps
Reid explains that one of the planned developments for Scholar Gateway is to expand its content offerings beyond scientific articles to include books and news. In addition, Wiley intends to improve its semantic search and develop new features focused on citations and repetitive tasks.
He says he is optimistic and sees a range of possibilities for applying the tool. “The question for us is prioritizing,” he reflects.
“We've got a lot of users, we want to be guided by what they're asking us and what they're doing. What problems is Scholar Gateway really solving? We want to focus our roadmap on that,” Reid emphasizes.
One avenue Wiley is beginning to explore is how Scholar Gateway can democratize access to scientific knowledge. “Think of a science journalist. In an ideal world, they would go and read lots of papers and check a story against the literature, but sadly that doesn't happen in reality because no one has time to do that. Can Scholar Gateway help serve that?” reflects Reid, excited by the possibility.
“I'm being very mission driven about it in terms of putting our very specialist, very high quality content in the hands of people who need it, but otherwise may not have been able to understand it.”