MCP Apps starts moving real work into the AI chat interface

Since its release in November 2024, the Model Context Protocol or MCP has provided a standard way for developers to connect AI assistants to external applications. Last week, the standard took a leap forward with the release of MCP Apps, a new way to embed applications directly into the chat. 

Even early on, the protocol was gaining traction. I recall a conversation with Jon Turow, a partner at Madrona Ventures, at AWS re:Invent a couple of weeks after MCP was released that underscored just how much curiosity it was already generating. He saw clear potential, even though the protocol still had real limitations at the time.

His observations proved prescient. MCP continued to evolve throughout last year as developers saw it as a way to solve a sticky application communication problem. Anthropic ultimately donated MCP to the Linux Foundation to ensure it would remain an industry standard.

While AI assistants like Claude, Gemini and ChatGPT can be tremendously helpful in reducing the amount of friction involved in getting work done, they still require a user to ask the question and then switch to one or more applications to do the task. MCP Apps has the potential to let you complete a set of tasks in one place without constantly switching.

It's not unlike the Slack value proposition a decade ago. While chat programs existed prior to Slack, it differentiated itself from the pack by using plug-ins for external enterprise applications like Salesforce or Workday without leaving Slack. This offers the same potential, but with greater intelligence.

Box CEO Aaron Levie, whose company is included in the initial group of supported apps, sees this as a natural progression for the protocol, one that helps extend the capabilities in a meaningful way beyond simple text. "Sometimes you want to be able to render an object or be able to create a custom interface or have something that somebody clicks on to kick off something, and it doesn't really just work as a text prompt. So I think MCP Apps is a necessary part of the evolution of MCP," Levie told FastForward.

Moving beyond simple connections

Levie is hardly alone in seeing MCP Apps as a natural next step for applications like Box, but Jay Anderson, an analyst at Moor Insights & Strategy, is thinking about what it means for the broader AI and software stack.

He says that while MCP initially solved a developer problem by making it easier to connect to external apps, MCP Apps solves a knowledge worker problem of users having to mash answers from the assistant into their workflow, rather than working within a smooth, integrated process. "So this integrated experience, I think, is a really big deal," he said.

This capability could help push agentic AI, enabling companies to build workflows involving multiple apps. Anderson sees this going a couple of ways. The first is built-in experiences inside the applications. "There are going to be these embedded agents that kind of drive an experience," he said. Those experiences will be built into applications like Slack, Google Workspace and Microsoft Office.

The other approach would involve more custom agents built one at a time to solve a particular problem. "So you're kind of seeing the market move towards some agents as part of an integrated experience and some agents as a bespoke experience. And those are starting to separate," he said. Regardless, he sees MCP Apps as the glue that connects both kinds of agents to these external apps.

Still an incomplete vision

But IDC analyst Arnal Dayaratna says MCP Apps is really just another step along the way, and an incomplete one at that. He believes realizing a broader agentic vision will require the ability to scale, something he is not convinced the protocol can achieve in its current state because it’s still essentially a per‑app connector.

What he means is that it doesn’t help the model choose the right endpoint — the specific system or account, such as a given Box or Dropbox workspace — or coordinate actions across these multiple apps from a single, unified context. 

Right now, he sees MCP Apps largely as one‑to‑one integrations between the protocol and individual applications like Figma or Slack, rather than something that can orchestrate workflows across multiple tools. “The protocol is very siloed and it focuses on individual technologies,” he says, and therein lies the problem. 

This is not trivial, and it means that it will likely take another big leap for this kind of cross-app integration to evolve. But MCP Apps is still a significant step forward, it clearly doesn’t completely fulfill a multi-application agentic vision. Even so, this ability to build workflows inside a chat interface that includes many of the productivity programs people are using already represents a meaningful advance.