Healthcare AI has a new three-letter acronym: MCP. It’s showing up on press releases, product pages, and social posts, often with claims that it solves systemic data problems in healthcare.
Some of that is true. Most of it isn’t.
Large language models are powerful but frozen in time. They can’t access real-time data or act in the world, like update a patient record or check a live provider schedule. Model Context Protocol (MCP), introduced by Anthropic in November 2024, addresses this directly. It’s an open standard that gives AI a consistent way to connect to external data sources and turn a static model into one that can retrieve information and take action.
Below, we walk through the myths, realities, and the messier middle ground in between.
MCP is Being Oversold. Here’s how.
Is the ability to connect to data the same as understanding it? And what happens when that data is wrong to begin with? These questions begin to expose where the hype separates from reality.
MCP is an interoperability layer that standardizes how AI connects to external data and tools. Before MCP, every new integration required custom-built connections. Now, MCP replaces that with a single protocol that any compliant system can use to enable AI to work with new data and access capabilities that weren’t part of its original training. And that’s genuinely useful, but stops short of claims being made that MCP reduces hallucinations, creates clinical-grade reliability, or can power care navigation out of the box.
Another misconception is that MCP is a replacement for traditional APIs. And this distinction is important. APIs are deterministic (i.e., Program A tells Program B to do exactly X). MCP is a structured transport layer. The semantic decision-making —which tool to call and when—happens at the AI model layer, on top of MCP. That flexibility is part of the appeal. It’s also where things go wrong.
The risk of bad routing decisions isn’t unique to MCP, as it’s a feature of any agentic AI set up. When the model is making decisions, you’re one bad inference away from the wrong tool being called, the wrong dataset being pulled, or a confident-sounding answer built on the wrong foundation. MCP is the plumbing, and moves water. It doesn’t filter it.
A Plain-English Guide to MCP.
Think of MCP like a USB-C port for AI. Just as USB-C gave the world a single connector for any device, MCP gives AI a single, standardized way to connect to any compliant data source or tool. Here’s what that means in practice:
- Establishes secure connections between AI and external datasets and tools
- Acts as a discovery layer (i.e., the model can ask “what tooling is available?” before it starts working)
- Translates data into a format the model can read and process
- Does all of this in real time
MCP’s three-part architecture—hosts (the AI application you interact with), clients (the connection manager inside the host), and servers (lightweight wrappers that advertise what each tool can do)—is what makes plug-and-play connectivity possible.
Connectivity Isn’t Really the Issue. Data is.
Connectivity is no longer healthcare’s problem. For many health systems, the hardest connectivity problems have been solved, as HL7, FHIR, and REST APIs have been moving data for years. MCP doesn’t solve what was already addressed, nor solve data quality. If your underlying data is messy, fragmented, or duplicative, MCP gives AI a faster, more confident way to ingest those errors. The garbage in, garbage out problem doesn’t disappear, but accelerates.
Take the management of your provider data. If three conflicting phone numbers exist across three different datasets, MCP will surface all of them. What the AI model does with that conflict—which it picks, or whether it fabricates a composite—is an LLM behavior. MCP has no built-in logic for data validation and no way to flag which source is authoritative.
The Gaps MCP Leaves Behind.
Any promise that MCP will “fix data silos” is selling you the pipe and calling it the water treatment plant. If the data on the other end is fragmented or poorly defined, you now have multiple broken sources of truth being called simultaneously by a model that has no way to know which one is right.
MCP won’t:
- Verify that data is accurate
- Tell the AI which source is authoritative
- Fix broken schemas or missing fields
- Ensure the data being retrieved is relevant to the question being asked
What you’re left with is a faster, shinier wrapper around the same messy data. And while accessibility gets solved, data integrity doesn’t. Over time, the problem compounds: dozens of poorly managed MCP servers mean higher token costs, more context bloat, and an AI that’s increasingly likely to grab the wrong tool or pull from the wrong dataset. More connections don’t make the answers better, but harder to trace.
Good Plumbing. Wrong Problem.
MCP accelerates whatever’s underneath it. If that’s fragmented scheduling logic built on stale decision trees, tribal knowledge, and 15,000 EMR configurations nobody can update fast enough, you haven’t solved the problem. Rather, you’ve automated it, and at greater scale and speed than any human ever could.
The question MCP can’t answer is: what should the AI find when it gets there? The answer is DexCare.
DexCare is the layer MCP was always assuming existed. Before AI connects, before a patient searches, before a staff member routes a call, DexCare digitizes and centralizes the clinical and operational logic that determines who gets seen, where, and when. Scheduling rules. Provider capacity. Care pathways. And even Sharon’s 20 years of expertise working at the front desk. All your data and institutional knowledge are governed in one place, not scattered across spreadsheets, EMR workarounds, and institutional memory.
DexCare is where your data gets smart. Every rule, every pathway, every hard-won clinical insight is captured, reconciled, and kept current so AI has something worth finding.
The health systems that win tomorrow are the ones that get access right today. DexCare is the growth engine that gets them there.
- 40% more appointments, same clinical resources
- 5 fewer days waiting for care
- 25% increase in net-new patients
