Your AI chat can now talk to the image archive

Anyone using an AI chat in a practice today spends a lot of time copying. A report from the archive into the chat, the summary back into the medical record, a question into the search field of a third program. The chat understands the question, but it sees nothing of what it is supposed to talk about. The work between the windows stays with the people.

That is not the fault of the chat, nor of the specialist software. What is missing is the connection between them. Until now, every program would have needed its own integration for every chat, and nobody built them.

What an MCP server is

MCP stands for Model Context Protocol. It is an open standard that Anthropic introduced in November 2024 and that the major AI assistants now support. It defines how an AI chat queries another piece of software and gives it tasks: which actions exist, what each one needs and what it returns.

The MCP server is the specialist software's side of the connection. It tells the chat what is possible, for example "find a patient's studies" or "read a report", and carries it out when the chat asks. The software stays as it is. It simply gets a second door, one you walk through with sentences rather than clicks.

The comparison that holds up best is a new practice assistant on her first day. She doesn't know the routines yet, but you can tell her what you need, and she knows where to look. That is exactly what the chat gets: not the whole archive, but the actions you use to work in it.

Six tasks for an image archive

What this means day to day is best shown with sentences that are spoken in a practice anyway. Each of them is currently a small search in a second program.

None of these tasks is spectacular on its own. The difference is that you no longer have to translate them into five clicks, and that they can be chained: search, summarise, pass on, in one sentence and one window.

Several servers, one conversation

A chat is not limited to one server. It can be connected at the same time to the image archive, the practice mailbox, the calendar or the practice system, each through its own MCP server. It decides for itself which one a question needs.

The example above shows why that makes the difference. A referral arrives by email, and the images that go with it are in the PACS. Today that means two programs and a search in each. With two connected servers it is one question: the chat opens the email, reads the attachment, finds the studies by name and date of birth and, once confirmed, files the referral alongside them.

Where the line is drawn

A chat that can work in the archive has to follow the same rules as a person who does. Three of them are not negotiable.

• The chat works with the permissions of the person using it. Anyone who may not see a study in the archive won't get it through the chat either.
• Reading happens straight away, changes only after confirmation. The chat proposes assigning, sending and anonymising; they are carried out only once a person agrees.
• Every task is logged with the person, the time and the result. What the chat did can be traced later, like any other change in the archive.

What matters when choosing

Every vendor now says its software "does AI". The more useful question is whether it can be operated by the chat you already have, or whether it forces a second one on you. Five questions help put it in context.

• Is there an MCP server, or only a built-in assistant that works exclusively inside this one program?
• Does the chat work with the permissions of the individual user, or with a shared account that sees everything?
• Which actions change data, and does the software ask for confirmation on each of them?
• Is there a log of what was triggered through the chat?
• Can you switch chats without rebuilding the integration? That is exactly what an open standard is for.

The last question is underestimated. AI models change every six months or so. An integration tied to one particular vendor ages along with it.

In Unomed PACS

Unomed PACS has an MCP server. It provides the actions from the table above: searching studies, creating viewer links, reading reports, preparing links that need no account, reviewing the inbox and anonymising studies. It works with the permissions of the logged-in user, asks before every change and logs every task. The archive is hosted in certified data centres in Switzerland.

The diagram below shows how it fits together. On the left are the places where questions are asked, in the middle the MCP servers, on the right what gets done. The PACS server is one of them; a server for the mailbox or the calendar can be added alongside it in the same chat. The same server also runs fixed workflows that need no question, such as pre-sorting the inbox overnight.