Athenahealth connects Claude to the medical records of more than 20% of US residents

A medical record holds a great deal of useful information, but even a simple fact can be hard for a person to find in it. The latest lab result may sit between a visit note, a list of prescriptions and cryptic abbreviations. Athenahealth offers a way out that developers will find familiar: connect the data to a language model via MCP and let patients ask questions in everyday words.

The company has launched AthenaPatient MCP, a server through which Claude gets access to selected sections of a medical record after the patient's explicit authorization. Athenahealth serves more than 20% of the US population, so this is not a lab demo but a potentially major channel of access to sensitive data.

What it looks like for the patient

The user signs in to their Athenahealth account, allows the connection and chooses which information can be shared with Claude. The model can then take into account lab results, vital signs, medications, allergies, vaccinations, diagnoses, visit history, procedures and treatment plans. The permission can be revoked.

A simple example: instead of searching through documents by hand, the patient asks, "What was my most recent red blood cell count?" Claude finds the relevant entry and returns the value from the record. But the boundary is drawn deliberately: the service helps retrieve and explain information that has already been recorded, and it does not prescribe treatment or replace a doctor. The question of why a value fell outside the normal range and what to do about it should go to a specialist.

The connection is currently available to users of the paid Claude Pro and Claude Max plans. Having a subscription opens nothing by itself: the patient must complete authorization separately and agree to the transfer of data.

Why MCP matters more here than the chatbot itself

For developers, the interesting part is not the question interface but the architectural pattern. MCP turns a medical system into a managed source of context: the model does not receive the entire database, but queries data on behalf of a specific user and within the rights granted.

This approach can be carried over to banks, insurers, corporate repositories and government services. However, a single "Allow access" button is not enough. Sensitive systems need narrow permission scopes, short-lived tokens, a log of every request, clear access revocation and a ban on retaining unnecessary context. The interface should also show where an answer came from: which document, which date and what status.

Athenahealth states that the terms of the agreement prohibit Anthropic from using information obtained through the integration to train models or improve products. That is an important promise, but a product team still has to design the system as if any unnecessary disclosure of data would become an incident. Broad permissions, hidden background requests and answers with no link to the primary record are especially dangerous.

What to check before a similar launch

The first question is whether the model really needs access to the entire entity or whether a few fields are enough. The second is whether the user can see and cancel each permission. The third is what happens if Claude makes a mistake, mixes up dates or confidently restates an outdated diagnosis.

The practical takeaway is simple: MCP speeds up integrating a model with closed systems, but it does not solve security automatically. For a product, value appears where the assistant removes routine searching and does not pretend to be a doctor. For a developer, the minimum set is targeted authorization, verifiable answer sources, a full audit of requests and a safe refusal when data is insufficient. These details will determine whether a medical assistant becomes a useful interface to the record or just another route for leaks.