
The Model Context Protocol makes it easy to give an AI agent access to tools and data. In healthcare, that ease is exactly why the questions matter. This guide covers the 9 things to settle before an agent touches patient data.
AI agents are already part of healthcare work. They summarise records, answer questions about lab trends, draft letters, prepare data for analysis and support programme teams. The Model Context Protocol (MCP) has become a common way to connect them to the tools and data they need: an MCP server exposes a set of tools, and any compatible AI application can discover and call them.
That is a real step forward. It means teams no longer have to build a custom integration for every assistant and every data source. It also means an agent can be given access to sensitive health data in an afternoon. Whether that access is safe, useful and defensible depends on decisions that are easy to skip when a demo is working.
This guide is written for product and engineering leads at digital health companies, hospital IT and innovation teams, pharma AI teams, and the compliance and security colleagues who need to approve what they build. It is organised in three parts: access, the data and the content the agent sees, and oversight over time.
Figure 1. The path from a person to health data through an AI application and an MCP server. Each numbered marker shows where one of the 9 questions in this guide applies: who the agent acts for, what it can see and do, how content is treated, what is logged, where data goes and whether the data is ready.
Figure 1 shows the path every request takes. The person asks, the AI application decides which tools to call, the MCP server carries out those calls, and the data sources answer. The 9 questions map to points along that path, and a weak answer at any one of them weakens the whole connection.
The simplest way to reduce risk is to reduce reach. An MCP server defines its tools, and each tool defines what data it returns. A tool called "get patient record" that returns everything is very different from one that returns the last five HbA1c results for a patient the user is already treating.
Healthcare already has a principle for this. Under HIPAA (the US Health Insurance Portability and Accountability Act), the minimum necessary standard asks organisations to limit uses, disclosures and requests of protected health information to the minimum needed for the purpose.6 Similar ideas run through data protection laws in Asia, Europe and Africa. Apply it to tool design: write down the task the agent supports, then expose only the patients, data types and fields that task needs.
Reading data and acting on it are different levels of risk. The OWASP (Open Worldwide Application Security Project) Top 10 for LLM Applications names "excessive agency" as a key risk, and expanded it in the 2025 edition because agents are increasingly given tools, permissions and autonomy.4 The MCP tools specification itself says there should always be a human in the loop with the ability to deny tool invocations.3
Figure 2. Four levels of agent capability. Each level adds risk and needs every control from the levels before it, plus its own. Most healthcare use cases today are well served at levels 1 to 3, where a person stays responsible for anything that changes a record or reaches a patient.
As Figure 2 shows, the question is not whether agents may act, but which actions, at what level and with whose approval. Separate read tools from write tools, make write tools draft rather than commit wherever possible, and keep anything irreversible or patient-facing behind a human decision.
An agent should never see more than the person using it is allowed to see. That sounds obvious, but it is easy to break: an MCP server connected to a database with a single service account gives every user of the agent the same, usually broad, access.
The MCP specification includes an authorization framework for HTTP-based connections. Authorization is optional in the protocol, but where it is supported it is built on OAuth 2.1.1 The MCP security guidance also explicitly forbids "token passthrough", where a server accepts a token from a client and passes it straight to a downstream system without checking it was issued for that server.2 For health data, treat authorization as required, tie every tool call to an identified user, and enforce that user's permissions at the point where data is retrieved, not only in the AI application.
Agents answer confidently from whatever they retrieve. If the data is inconsistent, the answer will be confidently wrong. This is the question most teams skip, because the agent appears to work in testing on a handful of clean records.
Figure 3. The same question asked of two data layers. With raw results in mixed units, the agent sees 7.2, 58 and 6.9 and either reports a false spike or drops a value. Harmonised to one unit using the published HbA1c master equation, the results are 55, 58 and 52 mmol/mol: a modest rise, then an improvement.
Figure 3 uses one of the most common examples. HbA1c (glycated haemoglobin) can be reported as a percentage or in mmol/mol, and the two are linked by a published conversion.8 An agent that does not know which unit each value uses will mislead the person asking. The same applies to tests with local names, reference ranges that differ by lab, and records that exist only as scanned PDFs the agent cannot read reliably. Before connecting an agent, make sure the data behind each tool is structured, harmonised and mapped to medical coding standards, so the agent is interpreting clinical facts rather than formatting quirks.
When an agent calls a tool, the result usually goes back into the model's context. That means health data may leave the system it lives in and travel to wherever the AI application and model run. Ask where the model is hosted, which region processes the data, whether prompts and outputs are retained or used for training, and which providers sit in between.
For many organisations, the answer has to satisfy local rules on in-country processing and cross-border transfer, as well as their own policies. Reducing what the agent retrieves (question 1) and returning pseudonymised data where identifiable data is not needed both reduce how much sensitive data travels at all.
Prompt injection is the first risk in the OWASP Top 10 for LLM Applications.4 It happens when text the model reads contains instructions that try to change what it does. In healthcare, agents read referral letters, patient-uploaded documents, messages and notes from many sources, so the content an agent retrieves can carry instructions too.
MCP adds a second channel. In April 2025, Invariant Labs described "tool poisoning" attacks, where malicious instructions are hidden in MCP tool descriptions that are visible to the AI model but not to the user.5 The MCP specification also says clients must treat tool annotations as untrusted unless they come from trusted servers.3 Practical defences include connecting only to MCP servers you trust and have reviewed, keeping tools narrow (so a manipulated request cannot do much), treating retrieved content as data rather than instructions, and keeping sensitive actions behind human approval.
If an agent retrieves the wrong patient's record or acts on a misread result, you need to know what happened. Healthcare already expects this: the HIPAA Security Rule requires mechanisms that record and examine activity in information systems that contain or use electronic protected health information.7 Agent activity should be no exception.
At a minimum, log which user the agent acted for, which tool was called, with which parameters, what data was returned (or a reference to it), and what the agent did next. Keep the logs where your security team can review them, and agree how long they are retained. Logs are also how you learn which tools are used, which fail and where the agent needs better data or clearer instructions.
A good human colleague says "I'm not certain" or "this result looks odd, let me check". An agent will not unless it is designed to. Build tools that return status and quality information alongside data: whether a value has been clinically reviewed, whether a unit was converted, whether a document could only be partly read. Then make sure the AI application surfaces this to the person rather than smoothing it away.
For anything clinically important, decide where clinically trained people review the output, and make that review easy. This is the same principle as human-in-the-loop review of data transformation: automation handles the volume, and people check what is flagged. Our 5 Levels of Human-in-the-Loop guide sets out what that looks like at each level of maturity.
Agents sit on top of several moving parts: the model, the AI application, the MCP server's tools, the data schemas behind them and the data itself. A change to any of them can change the agent's behaviour, sometimes in ways nobody notices until a wrong answer reaches a clinician.
Version your tools and schemas, keep a set of test questions with known correct answers (including tricky cases such as mixed units and negated findings), and run them whenever a model, tool or data source changes. The MCP specification allows servers to notify clients when the list of available tools changes;3 treat a changed tool as a change that needs review, not a silent update.
Table 1 brings the 9 questions together for your product, engineering, clinical and compliance teams.
Table 1. A checklist of the 9 questions and what a good answer looks like for each.
The thread running through all 9 questions: give the agent the least it needs, make sure what it reads is ready to be interpreted, and keep a person accountable for anything that matters.
Teams usually come to us with one of two questions. Some want to know whether their data and their plans are ready for agents at all: which use cases make sense, what access is appropriate and how to govern it. Others already have agents and need the data underneath them to be reliable. We support both.
JondaX has processed more than 500,000 files and 26 million data points at 99.9 per cent field-level accuracy, across more than 10 languages. The same engine sits behind the API and the MCP server, so whether your team integrates directly or through an agent, the data it receives has been through the same harmonisation and review. API documentation is available at jondax.redocly.app.
An MCP (Model Context Protocol) server is a standard way of giving AI assistants and agents access to tools and data. In healthcare, an MCP server might let an agent look up lab results, structure a clinical document or search a patient's history. The server decides which tools exist and what each one can do.
It can be, if access is limited to what the task needs, the agent acts with the permissions of a specific user, data stays where your rules require, every tool call is logged and sensitive actions need human approval. The risks come from broad access, unchecked actions, untrusted content steering the agent and data that is not ready to be interpreted.
Prompt injection is when text the agent reads, such as a document, a message or a tool description, contains instructions that try to change what the agent does.4,5 In healthcare, records and documents come from many sources, so an agent with access to patient data should treat their content as data, not instructions, and should not have permissions that let a manipulated request do harm.
For health data, yes. Authorization is optional in the MCP specification, but where it is implemented over HTTP it is based on OAuth 2.1,1 and the security guidance forbids passing tokens straight through to downstream systems.2 In practice, an MCP server that touches health data should know which user an agent is acting for and enforce that user's permissions.
Agents answer confidently from whatever they retrieve. If lab results are in mixed units, tests have local names and records are scanned PDFs, an agent can draw the wrong conclusion and present it as fact. Harmonised, structured data gives the agent something it can interpret consistently.
Yes. JondaX, from Jonda Health, has an MCP server alongside its API, so AI agents and assistants can work with structured, harmonised clinical data from labs, medical devices and clinical documents. Email hello@jonda.health for access.
About the author. Suhina Singh is the Founder & CEO of Jonda Health, a health data infrastructure company that turns fragmented clinical data, digital and non-digital, into structured, standardised information that clinicians and systems can actually use. Jonda Health is ISO 27001 certified and works with digital health companies, hospitals, laboratories, registries and life sciences teams. Get in touch at hello@jonda.health.