Over the past 18 months, the largest AI companies in the world have quietly settled on an approach to building the next generation of apps and services — an approach that would allow AI agents from any company to easily access information and tools across the internet in a standardized way. It’s a key step toward building a usable ecosystem of AI agents that might actually pay off some of the enormous investments these companies have made, and it all starts with three letters: MCP.
MCP, or Model Context Protocol, began as a passion project from two Anthropic employees, but since its creation in mid-2024, it’s been widely adopted by companies like OpenAI, Google, Microsoft, and Cursor. There are even hints that Apple will use MCP in its forthcoming AI-enabled version of Siri. There have been competitors to MCP, but so far it’s been a standards war without any real battle — MCP has quickly taken over the industry.
And now it’s official: This week, Anthropic is donating MCP to the Linux Foundation — and joining OpenAI, Google, Microsoft, AWS, Block, Bloomberg, and Cloudflare in establishing a new fund called the Agentic AI Foundation (AAIF), whose goal is to “advance open-source agentic AI.” The donation, and assigning a neutral body to govern MCP, will likely help supercharge its growth.
It’s also a move that should change up how AI systems operate as we know it. For AI companies, MCP is the new standard for how these systems should access apps, tools, and information — and by extension, how people use the internet.
A “ping-pong of intelligence.”
MCP essentially tells AI models which external tools, data sources, and workflows they’re able to access, then allows them to connect and perform tasks. When someone uses Claude to perform tasks in Slack, for example, MCP is what authorizes and establishes the connection between services. It’s what lets Claude redirect you to Slack and get notified once you’ve logged in. And it lets Slack tell Claude which tools, resources, and features it can access — “essentially a ‘show me what you’ve got,’” Conor Kelly, a product marketing manager for MCP at Anthropic, says.
From the user’s side, this simply means Slack and Claude can easily work together — a “ping-pong of intelligence,” as Anthropic CPO Mike Krieger puts the impact of MCP. When somebody prompts Claude to send a Slack message to a colleague, Claude knows that the Slack MCP server is connected, that a tool exists for sending messages, and that it can access that tool. Once it’s all set, Slack tells Claude that it happened successfully, then Claude tells the user. Message sent.
If you’re familiar with how computers generally worked before AI, this might all sound like a bunch of APIs — and you might recall that web apps and services opening their APIs to one another was the underpinning of the Web 2.0 era, and eventually the enormously lucrative explosion of mobile apps in the app store era. Moving users (and their money) from apps and websites to AI agents is one of the few ways AI companies can even begin to pay off their enormous investments. But AI agents need new kinds of APIs, and MCP seems like the standard those APIs will take. MCP’s webpage, aspirationally, likens it to the ubiquitous USB-C.
MCP started as a pet project by two Anthropic engineers, David Soria Parra and Justin Spahr-Summers. The initial goal wasn’t to build an industry-wide standard. The pair simply wanted Anthropic’s staff base to use Claude more in everyday work. They felt like something was missing in the chatbot: the ability, Soria Parra tells The Verge, to connect “to the outer world that you actually deeply care about, the things you interact with.” His initial name for the service was Claude Connect.
Other Anthropic employees, it turned out, agreed with them. In an October 2024 hackathon, virtually everyone used the protocol to build their projects — “It was this moment in the company when everybody was like, ‘Oh, there’s really something to this,’” Soria Parra says. He and Spahr-Summers got Krieger’s approval to develop a full-fledged open source project. They released it just before Thanksgiving — somewhat deliberately, Soria Parra says, so that people could take a break from family to check it out.
The protocol has gotten top billing on a San Francisco billboard.
Krieger says that initially, his “dream case” for MCP was getting just one other frontier lab to adopt it. But widespread adoption came fast. On March 19th, Microsoft announced it would support MCP in its Copilot Studio. One week later, OpenAI CEO Sam Altman posted that “people love MCP and we are excited to add support across our products.” Four days after that, Google CEO Sundar Pichai took a poll approach, posting, “To MCP or not to MCP, that’s the question. Lmk in comments.” The protocol has gotten top billing on a San Francisco billboard, and there are even hints of MCP support in beta versions of iOS, suggesting Apple’s false start on agentic Siri might be turned around by adopting the nascent standard.
MCP has caught on partly because its creators have spent so much time watching and learning what developers actually wanted from AI systems. It “encapsulated patterns that already existed at the time,” says Soria Parra. OpenAI uses MCP to underpin the ChatGPT apps it introduced earlier this year, such as Booking.com, Canva, Coursera, Expedia, Figma, Spotify, and Zillow, as well as to connect to services like Notion and HubSpot. Anthropic uses it to enable connections with Slack, Asana, Box, Square, Stripe, and many others.
At this point, MCP has effectively become a multi-company project. A group of “core maintainers” have an ongoing discussion on Discord and GitHub, and representatives from Google, Microsoft, OpenAI, and others periodically meet face-to-face, discussing ways to fix problems and improve the protocol, OpenAI’s Nick Cooper, technical lead for the company’s approach to MCP, tells The Verge.
But MCP’s links with Anthropic have also potentially held the standard back. Anthropic has always made the protocol open source, but until now, any improvements by other companies could potentially contribute to their competitor’s intellectual property — and in theory, Anthropic could still one day choose to lock it down. Giving it to the Linux Foundation removes those concerns.
Anthropic isn’t the only company handing over something to the Linux Foundation. Block is donating Goose, its open-source AI agent, and OpenAI is donating Agents.md, which describes codebases to agents. Put the donations from Block, OpenAI, and Anthropic together, and the story is “about more than just MCP,” says Jackie Brosamer, head of data and AI at Block. “Protocols are essentially ways for systems to talk to each other, and that’s the most important thing to standardize.”
MCP’s links with Anthropic have also potentially held the standard back.
Jim Zemlin, CEO of the Linux Foundation — the largest organization for open source and standards in the world — has been in the industry for more than two decades and has personally overseen the creation or expansion of a handful of new standards and platforms. But even he has been shocked by MCP’s grassroots growth.
“I’ve never seen anything like this,” Zemlin tells The Verge. “I can barely keep up with the number of inbound calls from organizations who want to be a part of this. Usually I’m trying to convince someone, or scratching and clawing. This is really the reverse.”
Josh Blyskal, who leads strategy and research for AEO (think: essentially SEO for AI) at Profound, believes MCP will “absolutely become a standard, especially in commerce,” and he says in a year, the way AI companies currently scrape websites will look “antiquated.”
Krieger says that although he thinks there is still a future for people accessing services directly in apps and browsers, “there’s something very powerful” about an internet where most interactions run through MCP. “These things can operate kind of at the speed of LLMs versus at the speed of people,” he says. “You can issue 10 queries in parallel, you can do a data deep dive, versus navigating the web as was built for humans.”
“Everybody was like, ‘Oh, there’s really something to this.’”
MCP is also gaining traction as the industry faces a major problem: Although agents are an almost necessary step to making AI profitable, most of them simply don’t work very well right now. There are valuable enterprise uses, because enterprise environments can be more tightly controlled and made predictable for the agents to operate in. But for a range of consumer tasks, they’re still slower and less reliable than simply using the web yourself. And part of the reason is that they’re often surfing a web made for humans, not machines, to pull information from.
MCP may have improved agents somewhat already, but AI companies hope if it’s adopted more widely, it could dramatically mitigate the hangups frustrating users: that agents lag, that they require too much coaxing in order to complete a task, that they sometimes fail to perform a task at all. The protocol allows systems to talk directly between each other, ideally making them faster, more accurate, and more successful. Instead of an “app store,” the dream is to one day create a marketplace of tools for agents to use MCP to connect with — and when it comes to e-commerce, some of those tools will allow agents to “shop” on a user’s behalf, picking between price and performance based on a user’s budget. Users wouldn’t even need to learn MCP’s name; they’d just realize one day that agents can do something like plan a full trip — from booking flights and hotels to adding calendar reminders — faster than they can.
That might be the vision, but jumping feet-first into a specific standard like MCP is still a big bet. Meta CEO Mark Zuckerberg has publicly said it was a mistake for the company to invest so much in HTML5 at the dawn of the mobile era instead of native apps, for example. And HTTP won out in the early days of the internet over other protocols, like Gopher, which was developed at the University of Minnesota in the early 1990s. “Anytime you pick technology, it’s an implicit futures contract, because you want that technology to have a big ecosystem, lots of support, tons of developers,” Zemlin says.
“Standards can come and go,” Soria Parra says. “I don’t know what the AI industry will look like five years from now, and I think nobody’s able to predict this. Of course I would hope that MCP is still around.” Whatever happens, though, he says, “at least we have set the [stage] for more open standards in the field.”
“If you’re an end user and you don’t know much about technology, you should never have to hear about MCP,” says Soria Parra. “It should just work … For you, the model just does its magic and solves the task at hand.”
The move could also help mitigate the security nightmares that agentic AI poses, like prompt injection. Now that Anthropic won’t have ownership over MCP, other companies especially well versed in security can make improvements to the protocol that they may have shied away from making before. “One thing I think is really important about having this be more of an open project is that there are aspects that we as Anthropic don’t hit first, but other enterprises do — so authentication and security are two,” Krieger says.
“When everybody is working collectively on a standard, to help improve the technology so that it’s more secure, so that it can do effective trusted payments, so that it has a better way of communicating in a secure fashion, that’s when the market really gets made,” Zemlin says.
