What Exactly Is a Tool in MCP?

If you have been following the Model Context Protocol conversation, you have likely noticed that “tools” are becoming the central concept.

But what exactly is a tool in this context?

In MCP, a tool is an executable function exposed by a server that a large language model can discover and invoke. It is not just an API endpoint. It is a structured, callable action designed to be controlled by a model.

Think of tools as standardized bridges between reasoning and execution.

They can:

  • Wrap external APIs such as FHIR endpoints

  • Execute system commands

  • Perform structured data transformations

  • Orchestrate entire workflows

  • Trigger write-back operations into clinical systems

Each tool is defined with a structured input schema. It may also include behavioral annotations, such as whether it is read-only or write-capable. These annotations provide context to the model about what the tool does and how it should be used.

Tools are what allow LLMs to move from generating text to performing actions.


The Human-in-the-Loop Nuance

MCP is designed to be model-controlled. However, one nuance is often overlooked.

The specification recommends that tool execution include a human-in-the-loop. How that is implemented, however, is entirely up to the client.

This is not a minor detail.

The MCP client developer defines what “human-in-the-loop” means in their application. It could mean:

  • Explicit approval before a write action

  • Conditional approval based on risk level

  • Silent execution for read-only queries

  • Escalation to a human operator for sensitive workflows

This design decision affects product behavior, security posture, compliance alignment, and user experience. In healthcare, it may determine whether a system is assistive or autonomous.

MCP provides the structure. The product team defines the governance.


Why Introduce an MCP Server at All?

A common question arises:

Why introduce the extra layer of an MCP server? Why not connect directly to tools or APIs?

It is a fair question.

Introducing an MCP server can feel like additional overhead. There is schema design, server configuration, tool exposure, and client orchestration.

But abstraction always introduces short-term complexity to unlock long-term flexibility.

Without a standardized layer like MCP:

  • Every integration becomes custom

  • Each model-tool pairing becomes tightly coupled

  • Changes to tools require reworking application logic

  • Scaling across multiple models becomes fragile

With MCP tools:

  • Interfaces are standardized

  • Tool discovery becomes dynamic

  • Models remain decoupled from backend systems

  • Integrations become reusable and scalable

In other words, MCP transforms integrations from one-off connections into infrastructure.


Learning from Healthcare’s Past

Healthcare technology has already demonstrated the cost of fragmented, closed systems.

EHR ecosystems evolved with tight coupling, proprietary interfaces, and limited interoperability. The result has been complexity, vendor lock-in, and slow innovation cycles.

MCP represents an opportunity to avoid repeating that pattern in the AI era.

Instead of embedding tool logic directly into model-specific code, MCP introduces a shared abstraction layer between LLMs and real-world systems. That abstraction becomes the foundation for scalability.


From Prompt to Action

When an MCP server is in place, the lifecycle of a prompt becomes structured:

  1. A user submits a request.

  2. The LLM reasons about the request.

  3. The model identifies relevant tools exposed by the MCP server.

  4. The client determines whether and how the tool should be executed.

  5. The tool executes and returns structured output.

  6. The LLM incorporates the result into its response.

This structured orchestration is what enables real-world AI applications grounded in systems like FHIR servers, scheduling platforms, and clinical databases.


Watch the Architecture in Action

If you prefer a visual walkthrough of how MCP servers, clients, and tools interact, we’ve recorded a short explainer that breaks this down step by step.

▶ Watch the video below to see:

  • What MCP is and the problem it is designed to solve

  • The lifecycle of a prompt when an MCP server is involved

  • A real-world example of MCP interacting with a FHIR server

[Introduction to MCP Servers on MeldRx (ft. Dr. Pawan Jindal)]


Why This Matters for Healthcare

In healthcare, tools are not just convenience functions. They may:

  • Retrieve patient data

  • Submit prior authorizations

  • Write structured documentation

  • Trigger care coordination workflows

The difference between a chat interface and a clinical copilot lies in its ability to invoke tools safely and predictably.

MCP provides the framework to make that possible.


Continuing the Conversation

We recently recorded a video explaining the MCP server, client, and tool architecture in detail, including:

  • What MCP is and the problem it is designed to solve

  • The lifecycle of a prompt when an MCP server is involved

  • Real-world examples of MCP interacting with a FHIR server

If you are building at the intersection of AI, healthcare, and interoperability, these architectural decisions are not theoretical. They define how scalable and secure your system will be.

At Prompt Opinion, we are actively collaborating with teams exploring MCP-enabled architectures. If you are working in this space, we would welcome the opportunity to connect and exchange ideas.


Previous
Previous

FHIR-ify Your Data. LLM-ify Your Apps. MCP-ify Your LLM.

Next
Next

Let’s Talk About the Other Side of AI in Healthcare