MCP + Claude: An AI Operations Assistant Without PHI Risk

Every ABA operations leader has the same Monday-morning question: "Where are we?" How many referrals are sitting in Insurance? Which location has the biggest fax backlog? How many RBTs still owe a training module before the end of the month? Answering it usually means opening four screens, exporting two spreadsheets, and asking a coordinator to check a third thing.

Large language models are very good at answering exactly this kind of question — if they can see the data. The problem in healthcare is that "the data" is full of protected health information, and pasting it into a chat window is not an option. TargetFlo is building an MCP server for Claude that solves this by exposing only operational aggregates, never the underlying records. This post explains what that means, how it works, and what you will be able to ask.

What is an MCP server for clinical operations?

MCP (Model Context Protocol) is an open standard that lets an AI assistant such as Claude connect to external tools and data sources through a defined, permissioned interface. An MCP server for clinical operations is a small service that sits in front of an operations platform and answers a fixed set of questions — stage counts, task summaries, queue status — that the assistant can call as tools.

The important word is fixed. The assistant does not get a database connection. It gets a menu of aggregate queries that your organization has approved, each returning numbers and labels rather than records. TargetFlo's planned AI & MCP Assistant follows this design on top of the same metrics that power the live platform dashboard.

The PHI boundary: what the assistant can and cannot see

The value of an AI operations assistant depends entirely on where you draw the line. TargetFlo draws it at the aggregate.

Because the server returns counts and summaries only, the assistant can be helpful to a regional director without being a new place where PHI lives. The safety commitment is simple: AI access is scoped to operational aggregates. PHI never leaves your controlled environment.

This is the same principle behind TargetFlo's approach to HIPAA-aware architecture: separate schemas for organization data, clinical data, and operational metrics, so a tool that needs one does not automatically get the others. We cover the broader philosophy in What "HIPAA-Aware" Software Means for ABA Centers.

Example prompts for an ABA operations assistant

Here is the kind of conversation the MCP server is designed to support. Every one of these can be answered from aggregates alone.

• "How many referrals entered the Insurance stage this month across all locations?"
• "Which location has the most cards older than 14 days in Pending Documents?"
• "Compare enrollments per month for the last two quarters, by referral source."
• "How many faxes are in triage right now, and how many are older than 24 hours?"
• "What percentage of BCBAs have completed the updated supervision course?"
• "List the top three closure reasons for referrals closed in the last 90 days."
• "How many tasks are overdue, grouped by team?"
• "Summarize this week's pipeline movement: cards advanced, cards closed, cards added."

Notice what is missing. There is no "show me the Ramirez referral" and no "read me the last fax from Dr. Patel." Those questions belong inside the authenticated application, where role-based access and audit logging apply, not in an assistant.

Why aggregates are enough for most operational questions

It is tempting to assume an assistant needs record-level access to be useful. In practice, the questions that reach an operations leader are almost always about volume, velocity, and bottlenecks:

• Volume: how many referrals, faxes, tasks, or courses are in a given state.
• Velocity: how long items sit in each state and whether that is improving.
• Bottlenecks: which location, stage, or team is accumulating the backlog.

All three are answerable from counts and averages. When a leader does need a specific record, the right next step is a link into the intake pipeline or the task board, where the full record is available under normal permissions.

How the MCP server fits the TargetFlo architecture

TargetFlo already computes dashboard KPIs as lazy-loaded metric cards: stage counts, average age in stage, eligibility verification rate, overdue tasks, training completion, and more. Those metrics are calculated inside the platform from the clinical data, and only the results are stored in the operational metrics layer.

The planned MCP server reads from that metrics layer, not from the clinical schema. Each tool exposed to Claude maps to one metric family, accepts filters such as location and date range, and returns structured numbers. Requests are authenticated per organization, logged, and rate-limited. There is no free-text query path to the database.

The result is an assistant that inherits the boundary the platform already enforces. If a metric is not on the dashboard, it is not available to the assistant either.

What this looks like in practice

Consider a three-location center preparing for a weekly operations review. Today, the director asks each intake lead for numbers, pastes them into a slide, and notices on Friday that one location's fax backlog doubled on Tuesday.

With an MCP-connected assistant, the director asks Claude for the week's pipeline movement, the fax triage age by location, and overdue tasks by team, and gets a summary in a minute. The conversation is about what to do, not about gathering data. And because every answer is an aggregate, the transcript can be shared with the leadership team without a compliance review.

What is live now and what is coming

To be clear about status:

• Live today: dashboard KPIs, pipeline stage counts, task summaries, course completion tracking, fax queue status, and AI-assisted OCR through Gemini-powered integrations.
• On the roadmap: the MCP server that connects those aggregates to Claude. TargetFlo is building it on the metrics layer described above, and the safety boundary is a design constraint, not a later add-on.

If you are evaluating AI for operations, the question to ask any vendor is not "can it answer my questions?" but "what exactly can it read?" For TargetFlo, the answer is operational aggregates and nothing else. For the metrics that assistant will be reading, see 10 Dashboard KPIs Every ABA Operations Leader Should Track.