How AI Talks to Your Tools

The Consult/Referral Model

Every day in clinical practice, you send structured requests to specialists. A referral has a specific format: patient demographics, clinical question, relevant history, and what you need back. The specialist has a defined scope of practice, processes your request, and returns a structured response.

AI integrations work exactly the same way. When ChatGPT searches the web or Claude reads a PDF, there's a structured request going out, a specialized service processing it, and a structured response coming back. The AI model itself doesn't "know" how to search the web any more than you personally run the MRI machine—it knows how to write the order and interpret the results.

This is the mental model that will carry you through the rest of this module: APIs are the referral system, tools and skills are the specialists, and MCP is the universal referral form.

APIs—The Referral

An API (Application Programming Interface) is a structured way for one piece of software to ask another for something specific. It's the referral form of the software world: a defined format, a specific recipient, and an expected response.

You already use dozens of systems that rely on APIs every day:

• Your EHR pulling lab results—the EHR sends a structured request to the lab system's API, which returns results in a standard format
• E-prescribing—your prescription order hits the pharmacy's API, which confirms receipt and flags interactions
• Insurance eligibility checks—a request goes to the payer's API, which returns coverage details in real time
• Appointment scheduling—when a patient books online, the scheduling widget talks to your practice management system's API

APIs have been around for decades. What's new isn't the concept—it's that AI models can now decide which APIs to call, compose the right request, and interpret the response without a human writing the code each time.

Tools and Skills—The Specialists on Call

Once AI can make API calls, it needs to know which specialists are available and how to consult them. These packaged capabilities go by different names depending on the platform—tools, skills, plugins, extensions—but they all work the same way: a defined scope of practice, a declaration of what information is needed, and a structured response format.

Here are real examples of tools AI assistants can use today:

• Search PubMed for recent evidence on a clinical question
• Read and summarize PDFs—that 47-page prior authorization document, for instance
• Query an EHR for patient history (in systems that support it)
• Send messages through Slack, email, or other communication platforms
• Browse the web for drug pricing, guideline updates, or formulary information
• Manage calendars—check availability, schedule meetings, send invites
• Run code—execute Python scripts, analyze data, generate charts
• Search and edit files on your local computer

Think of each tool like a specialist with a defined scope of practice. A PubMed search tool declares: "I can search the medical literature. Give me a query and optional filters, and I'll return relevant articles with abstracts." The AI model reads that declaration, decides when to use it, and interprets the results—just like a PCP coordinating care across multiple specialists.

Skills: The Evolving Frontier

The concept of what a "tool" is keeps expanding. The latest evolution is skills—higher-level capabilities that chain together multiple tools, remember context, and follow complex workflows. If a tool is a specialist you can consult, a skill is more like a protocol or care pathway: a defined sequence of steps that coordinates multiple specialists toward a specific outcome.

For example, a "literature review" skill might chain together PubMed search, PDF retrieval, summarization, and citation formatting—all triggered by a single request. A "patient prep" skill could pull the schedule, review recent notes, check pending labs, and generate a pre-visit summary. The AI isn't just calling one tool; it's orchestrating a workflow.

Skills are evolving fast across platforms:

• Claude supports custom skills through its coding tools, letting users define reusable workflows that persist across sessions
• ChatGPT offers Custom GPTs that bundle instructions, knowledge, and tool access into shareable packages
• Open-source agents like OpenClaw have skill marketplaces where community members publish and share capabilities
• Enterprise platforms are building internal skill libraries tailored to their organization's specific tools and data

The terminology is still settling—what one platform calls a "skill," another calls a "workflow," "agent," or "custom GPT." The underlying concept is the same: packaging multi-step, multi-tool capabilities into reusable units that an AI can invoke. This is where AI stops being a chatbot and starts acting more like a capable assistant.

Tool Ecosystems and Marketplaces

As tools and skills become more useful, platforms have started building marketplaces where developers can publish them for others to use. OpenAI has the GPT Store. Various open-source projects have their own plugin registries and skill directories. This creates enormous potential—and, as we'll explore in the OpenClaw module, enormous risk. A tool marketplace is only as trustworthy as its vetting process, and many of these ecosystems are moving faster than their security reviews.

MCP—The Universal Referral Form

Here's the problem MCP solves. Before MCP, every AI platform had its own proprietary way of connecting to tools. Building an EHR integration for Claude was completely different from building one for ChatGPT, which was completely different from building one for Gemini. If you wanted your tool to work everywhere, you had to build and maintain separate integrations for each platform.

Sound familiar? It's the same problem healthcare faced before FHIR (Fast Healthcare Interoperability Resources). Every hospital system had its own data format. Transferring a patient record meant custom interfaces between every pair of systems. FHIR said: here's one standard format for exchanging health data, and now everyone speaks the same language.

MCP (Model Context Protocol) does the same thing for AI tools. Published by Anthropic in late 2024 as an open standard, MCP defines a single way to package a tool so that any AI model can use it. Build your PubMed integration once using MCP, and it works with Claude, ChatGPT, Gemini, and any other platform that supports the protocol.

MCP Servers in the Wild

MCP has moved well past the whitepaper stage. Real MCP servers are running in production today, and the ecosystem is growing fast. Here's a sampling of what's available:

• PubMed—search and retrieve medical literature from the National Library of Medicine
• File systems—read, write, and search files on your local computer
• Slack and Gmail—search messages, read threads, send drafts, manage channels
• Google Calendar—view events, find free time, create and update appointments
• GitHub—manage code repositories, review pull requests, track issues
• Web browsing—navigate websites, take screenshots, extract content from pages
• Database access—query PostgreSQL, SQLite, and other databases directly
• Hugging Face—search AI models, datasets, and research papers

Notice the range. Some of these tools are read-only (searching PubMed), while others can take actions (sending emails, writing files, modifying databases). That distinction matters enormously for safety, and it's one you should keep in mind as AI tools get more capable.

How MCP Actually Works

The architecture is straightforward. An MCP server wraps around an existing service (like PubMed or your calendar) and exposes it in a standardized format. An MCP client is built into the AI application (Claude Desktop, for example) and knows how to discover and communicate with MCP servers. When you ask Claude to search PubMed:

1. Claude recognizes that answering your question requires the PubMed tool
2. The MCP client sends a standardized request to the PubMed MCP server
3. The MCP server translates that into a real API call to the National Library of Medicine
4. Results come back through the same standardized path
5. Claude interprets the results and gives you a useful answer

The key insight: the AI model never talks directly to PubMed. The MCP server is a middleman that translates between the AI world and the service world. This creates a natural place for access controls, audit logs, and security checks—though as we'll see in the OpenClaw module, not everyone implements those safeguards.

Where MCP Is Heading

The MCP ecosystem is evolving rapidly. Active developments include:

• MCP registries—centralized directories where you can discover and install MCP servers, similar to app stores
• Authentication standards—OAuth-based flows so MCP servers can securely access services on your behalf without storing raw credentials
• Enterprise deployments—organizations packaging internal tools (EHR access, billing systems, policy databases) as MCP servers for their staff's AI assistants
• EHR integrations—early-stage MCP servers connecting AI to clinical systems like Elation, Epic, and Cerner, though these are still largely experimental
• Cross-platform adoption—Google, Microsoft, and other major players have announced MCP support, validating it as an industry standard rather than a single-vendor project

This is infrastructure being built right now. The decisions being made about authentication, permissions, and audit trails in MCP today will shape how AI interacts with clinical systems for years to come.

Why This Matters for Clinicians

You might be thinking: "I don't need to know how the plumbing works. I just need to know if the tool is reliable." That's fair—and it's the same reasoning you apply to most technology in practice. You don't need to understand TCP/IP to use your EHR.

But there are three reasons this particular plumbing is worth understanding right now:

1. Actions Are Different from Answers

An AI that gives you a wrong answer is one kind of problem. An AI that does the wrong thing—sends the wrong message, modifies the wrong record, schedules the wrong appointment—is a fundamentally different kind of problem. As AI tools gain the ability to take actions through APIs, the stakes of errors change. Understanding that your AI assistant has tools that can read your calendar versus tools that can write to it changes how much supervision you should apply.

2. Data Flows Where Tools Flow

Every tool connection is a data pathway. When an AI assistant can access your email, your calendar, and your EHR, information from all three can flow through the AI model. This has profound implications for PHI. A question like "What did my patient with diabetes email about last week?" requires the AI to pull data from both clinical and communication systems—and the privacy considerations multiply.

3. The Ecosystem Is Growing Fast

MCP means the tools your AI can access will expand rapidly. Third parties can build integrations without needing permission from the AI vendor. This creates opportunity (a vibrant ecosystem of useful tools) and risk (a vibrant ecosystem of poorly vetted or actively malicious tools). As the next module will make painfully clear, tool marketplaces can become attack surfaces.

Further Reading