MCP Apps and the Healthcare Ecosystem That's About to Explode
AI Stopped Being a Feature. It Became the Environment
Over the last few weeks, something shifted in AI that’s easy to miss if you’re only scanning announcements. It wasn’t just another model release or benchmark bump. It was the feeling that AI quietly stopped being a feature and started becoming the place where software runs. That matters a lot for healthcare.
On the surface, the news looks familiar: Anthropic rolled out a wave of healthcare connectors and published what they call MCP apps small applications that run directly inside the Claude interface. OpenAI is clearly heading in a similar direction with ChatGPT. The obvious question people keep asking is which model is “better.” The more important question is what people can actually do with these environments now that they exist.
I stopped speculating and started building.
I took one of my own tools, Smart Health Connect. It was never meant to be a full-blown platform just a lightweight PHR and SMART-on-FHIR app. With standard Claude connectors and MCP-style tooling, I was able to spin up small apps directly inside the AI interface that let me explore visualizations and workflows without standing up a separate UI or backend. No big architecture decisions. No long setup. Just capability, right where the reasoning happens.
That’s the part that clicked for me. When apps live inside the AI environment, you’re no longer designing “a product” first. You’re designing a capability that can be composed, reused, and iterated on quickly. Once that’s true, the ecosystem explodes. There are easily a million use cases people can build with the same standard tools, even if the underlying data is imperfect.
And yes, the data is imperfect. Provider directories are messy. NPI data is barely usable out of the box. None of that is new. What is new is how little friction there is now between “this data is ugly” and “this is still useful.” I built a simple provider search MCP app that lets you see which providers are available in your area by specialty. It’s not complete. It’s not canonical. It’s still immediately more practical than most of the tools people rely on today. That’s a massive opening for companies willing to focus on usefulness instead of pretending the data problem is already solved.
Where this really starts to matter, though, is on the consumer side.
I’ve written before about the opportunity for patients to have something like a personal healthcare assistant one that handles most of the questions people currently Google, text their friends about, or bounce between providers to answer. This time, I actually put it to the test with my own data.
I didn’t get everything. Fragmentation is real. EMRs are still silos. But once I figured out where my data lived, I was able to pull it down through a mix of direct connections and free community tools that let anyone extract their Epic data in FHIR format, notes included. I also logged into a very niche primary care EHR my doctor uses and downloaded three years of labs and wellness visits.
For the first time, I had a true longitudinal record. Not a snapshot. Not a portal view frozen in time. I could look at my lipid trends across multiple providers over multiple years and actually reason about whether starting a statin made sense. I did the same exercise for my kids and for my wife.
Here’s the part that still feels uncomfortable to say out loud: I have better visualizations, better context, and better research on my family’s health now than I’ve gotten from any individual provider system, patient portal, or health app over the last three years. That’s not an indictment of clinicians. It’s an indictment of tools that were never designed for longitudinal reasoning.
This is why I think the opportunity around MCP-style apps whether from Claude or ChatGPT is so large. We should be innovating inside the AI ecosystem, not just bolting AI onto existing point solutions. When I open some patient apps today, there isn’t even a way to refresh the data. You import records once, admire the interface, and then it quietly goes stale. If the system can’t pull new data and update its recommendations, it’s not really helping anyone make decisions.
That’s also why I’ve been underwhelmed by the pace of change in products like Apple Health. I’m still genuinely curious what happens when ChatGPT Health becomes broadly available I’m on the waitlist like everyone else but at the moment, Anthropic feels clearly out front in terms of healthcare-specific tooling and published FHIR skills. And honestly, those skills are very cool.
All of this feeds into why I’m building a new platform called FHIR Builders. It’s not because the world needs another platform. It’s because too many people are reinventing the same plumbing over and over again. The idea is to give the FHIR IQ community a place where they can build apps on top of shared infrastructure, using standard AI skills, and surface those capabilities as MCP-style apps when it makes sense.
There’s no shortage of ideas. Scheduling alone something I spent time on recently remains deeply broken because it’s tightly coupled to individual EHR systems, provider inboxes, and slot logic. These are solvable problems, but not if every team has to start from scratch. My hope is to create an environment where experimentation is cheap, real use cases drive design, and people aren’t stuck demoing prototypes that never escape the sandbox.
Tooling finally makes this plausible. I’m spending most of my time right now in Claude Code, both inside VS Code and via the web and CLI. Deployment has become boring in the best way possible. You can get something live, buy a domain, and move on. With the latest Claude Opus releases, newer OpenAI coding models, and a growing set of MCP tools and skills, the cycle time for real innovation is collapsing.
What hasn’t collapsed is the difficulty of deploying something real in healthcare. Production is still hard. Real users are still messy. Regulations still matter. That’s exactly why shared infrastructure and communities matter more than ever.
I’m optimistic but not in a naive way. This is one of those moments where costly optimism is everywhere, and a lot of demos will never turn into anything useful. But the underlying shift is real. AI isn’t just helping healthcare software anymore. It’s becoming the environment where healthcare software lives.
If we get the plumbing right, the rest finally has a chance to follow.