MedaSystems launches beta MCP server for expanded access program data

"How many requests were approved, treated or withdrawn this year, by site?" It is one of five sample questions MedaSystems uses to present its new MCP server, launched in beta on 19 October. The company, a SaaS platform that pharmaceutical companies use to manage expanded access programs (known in Spain as compassionate use), now lets users run that kind of query from Claude, ChatGPT or Microsoft Copilot without exporting the data. It announced the release in a press release distributed by Business Wire.

The technology is not new at this point; the sector is. Expanded access programs, post-trial access programs and investigator-initiated studies (IIS) handle information about patients treated with drugs that are not yet approved. Few areas demand as much data governance.

What the server actually does

According to the full text of the announcement, the server is read-only: it does not create or modify records. Within that limit it covers a lot of ground:

Program-level and patient-level queries across cohorts, sites, investigators, treatment timelines and drug supply.
Structured outputs: listings, frequency tables and descriptive statistics.
Quality checks that detect missing values, duplicates, outliers and inconsistent terminology.
Scheduled reports that regenerate automatically from current data.

Among the examples the company offers are listing each dose reduction with the treatment cycle in which it occurred, summarising biomarker status for treated patients while flagging where it is unknown, or drafting a case narrative for every patient treated for more than ten cycles.

Permissions travel with the user

What interests us most about the announcement is access control. The server inherits the single sign-on, role-based permissions and data governance customers already have configured in MedaSystems, and each user authenticates with their usual credentials. The rule is simple: if someone cannot see a record in the platform, neither can their assistant. Each customer gets a dedicated connection, the data stays in MedaSystems and the company says no patient data is written to logs.

Responses are also traceable: they indicate which records they come from and whether the information is complete. In a regulated environment, being able to audit where a figure comes from matters as much as the figure itself.

"We built the server so the platform's existing controls carry over unchanged," says Jeff Thibeault, Chief Technology Officer. Matt Rosen, CEO, frames it from the customer side: their users already keep their program data in MedaSystems and can now work with it in the AI tools they are adopting, without exporting it or giving up the controls they depend on.

Why it matters

We have seen many MCP servers that expose an API as is and rely on a token with broad permissions for security. The problem with that approach is that the assistant ends up seeing more than the person using it. The design MedaSystems describes flips that logic: the user's identity defines the perimeter and the model operates inside it.

Making the first version read-only also makes sense. Writing to a clinical record from a conversational assistant raises questions about validation, signatures and accountability that still have no standard answer. Starting with queries, reports and quality checks delivers value without touching those issues.

Who it is useful for

The audience is very specific: clinical operations, medical affairs and pharmacovigilance teams at companies that already run their programs in MedaSystems. For them, the practical change is to stop asking an analyst for a report, or exporting to a spreadsheet, and ask the question in the assistant they already use every day.

The beta is available today to all customers, who need to request it through their account team; MedaSystems handles the setup. No pricing has been published, and the general availability date will depend on what customers contribute during the beta.

For those of us who build MCP integrations, the case is a useful reference: connecting is the easy part, and the real work is making sure permissions, auditing and data minimisation reach the model intact. We will have to see how it performs beyond the beta, but the approach is the right one.