Claude Healthcare quick impressions

Clunky but can see the future

This isn’t a press release about “the future of medicine.” This is a report from the trenches of the “last mile” of healthcare data where the APIs meet the reality of forgotten passwords and legacy EHRs.

At the JPM conference last week, Anthropic essentially tried to turn Claude into a plug-and-play health analyst. They rolled out connectors for everything: CMS databases, NPI directories, ClinicalTrials.gov, and a massive pipeline via HealthEx Connector (powered by Fasten Health) to pull in medical records.

On paper, it’s the holy grail. In practice? It’s a reminder of how broken the “plumbing” of healthcare still is.

The “15-Click” Nightmare

I’ve been living in these new integrations for the last few days, and the experience is, frankly, exhausting.

The promise is a “straightforward” connection to your health system. The reality is a gauntlet of timeouts. If you’re like 99% of patients and don’t have your portal password memorized, you’re stuck in a loop of resets. Every time the connection refreshes, the Claude/HealthEx pipeline seems to lose its mind.

At one point, I had to click through 16 separate resource prompts just to get the data moving. If we expect a consumer or even a motivated patient to sit through that, we’ve already lost. It’s not a “connector”; it’s a part-time job.

The Interoperability Lie

Here’s the bigger problem: Interoperability is still a gated community.

I pulled data from UPMC and AHN via Epic (after logging in individually not through RLS). It worked, but it was hollow. Why? Because my primary care provider uses a custom EHR called Medent. Medent isn’t on the “cool kids” list of connected systems.

This creates a massive data “blind spot.” All my labs and testing from the last several years the stuff that actually matters for longitudinal health stayed locked in Medent. Even with USCDI standards, if the systems aren’t talking, the “AI Summary” you get is effectively meaningless. An AI is only as smart as the data it can actually see.

The Breakthrough: The “DIY PHR”

Despite the friction, I hit a “lightbulb” moment.

Since the automated pipes failed me, I manually exported my lab results from Medent and dropped the files directly into Claude. That is where the game changed.

Within seconds, Claude did what no patient portal has ever done effectively:

• It visualized my lipids over a three-year span.

• It highlighted the exact moments I needed to take action.

• It gave me a shareable link that I could send to my spouse and my doctor.

We’ve spent decades trying to build the “Perfect Patient Portal.” It turns out the best PHR (Personal Health Record) might not be a portal at all it’s just a smart LLM with a good UI and a manual upload button. I stopped being a “user” of a health system and became a “manager” of my own data.

The Builder’s Opportunity

For the app developers and product managers reading this: The opportunity isn’t in the pipes; it’s in the glue.

The “Record Locator Services” (RLS) aren’t working yet. The automated connections are fragile. If you can build the bridge that handles the “clunky” stuff the manual uploads, the OCR for PDF labs, the background syncing that doesn’t time out you will win.

Claude has the “brain” to be a world-class health assistant, but right now, it’s a brain without hands. It depends entirely on the user’s imagination and patience to feed it.

I’m still on the waitlist for the ChatGPT Health equivalent, and I’m eager to see if they’ve solved the “15-click” problem. But for now, if you’re willing to do the manual labor, the era of the “Personal Health Analyst” is officially here. It’s just a lot more manual than the JPM slides suggested.