AI & MCP Assistant
Coming soonAn AI operations assistant for your ABA center — without PHI risk
TargetFlo is building an MCP server that connects Claude to non-PHI operational data: pipeline counts, task backlog, fax queue status, and training completion. Ask in plain language. Get aggregates, never a client record.
How many referrals entered the Insurance stage this month, by location?
September so far: Raleigh 23, Charlotte 17, Durham 11 — 51 total, up 12% from August. Charlotte's median time in Insurance is 2.1 days.
Which location has the oldest faxes waiting in triage?
Durham has 4 documents older than 48 hours; the oldest is 3 days. Raleigh and Charlotte have none older than 24 hours.
What percentage of RBTs completed the HIPAA refresher?
87% organization-wide (61 of 70). Durham is at 74%, the lowest — 5 RBTs outstanding.
What is the TargetFlo AI & MCP Assistant?
The TargetFlo AI & MCP Assistant is a planned Model Context Protocol server that gives Claude read-only access to TargetFlo's operational aggregates — dashboard metrics, pipeline stage counts by location, task summaries, fax queue status, and course completion rates. It is an AI operations assistant for healthcare that uses non-PHI data only: client names, dates of birth, diagnoses, and insurance IDs are never available to the model. It builds on the live AI features described on the AI integrations page and the isolation described on the Security page.
Example prompts
Questions an operations leader could ask Claude
Every one of these is answerable from counts and rates. None requires knowing who a client is.
- 01
“How many referrals entered the Insurance stage this month, by location?”
- 02
“Which location has the most faxes waiting in triage older than 48 hours?”
- 03
“What percentage of RBTs have completed the HIPAA refresher course?”
- 04
“How many tasks are overdue for the intake team in Charlotte?”
- 05
“Compare Pending Authorization counts this month versus last month.”
- 06
“What is the median days-in-stage for Diagnostics across all sites?”
- 07
“How many referrals were closed as 'not accepted' this quarter?”
- 08
“Which courses have the lowest completion rate among new hires?”
- 09
“How many eligibility checks failed with 'member not found' this week?”
- 10
“Summarize this week's intake pipeline movement for the leadership meeting.”
How many referrals entered the Insurance stage this month, by location?
September so far: Raleigh 23, Charlotte 17, Durham 11 — 51 total, up 12% from August. Charlotte's median time in Insurance is 2.1 days.
Which location has the oldest faxes waiting in triage?
Durham has 4 documents older than 48 hours; the oldest is 3 days. Raleigh and Charlotte have none older than 24 hours.
What percentage of RBTs completed the HIPAA refresher?
87% organization-wide (61 of 70). Durham is at 74%, the lowest — 5 RBTs outstanding.
Data the assistant can use
Six operational sources, all aggregates
Dashboard metrics
The same KPIs leadership sees on the TargetFlo dashboard, queryable in plain language.
Pipeline counts by location
Cards per stage, per site, per period — with movement between stages over time.
Task summaries
Open, overdue, and completed task counts by team, role, and priority. Never the task body.
Fax queue status
How many documents are waiting, how old they are, and how many were linked today.
Course completion rates
Completion and certificate rates by course, role, and location for compliance reporting.
Eligibility outcomes
Counts of checks run and outcome categories — active, inactive, not found — with no member data.
Boundaries
What the assistant can and cannot see
The line is the PHI schema. Everything on the left is operational. Everything on the right is isolated and unreachable by the MCP server.
| Can see | Cannot see |
|---|---|
| Pipeline stage counts by location and time period | Client names, dates of birth, addresses |
| Median and average days-in-stage | Diagnosis codes or clinical assessment results |
| Fax queue size, age, and triage status counts | Fax contents, extracted fields, or document images |
| Task counts by status, owner role, and due date | Task titles or descriptions that reference a client |
| Course completion rates and certificate counts | Individual employee test scores tied to a name |
| Eligibility check volume and outcome categories | Member IDs, payer responses, or coverage details |
| Org-level KPIs already on the dashboard | Anything stored in the PHI schema |
Safety
AI access is scoped to operational aggregates. PHI never leaves your controlled environment.
The assistant is useful precisely because it is limited. It answers the questions a leadership meeting asks, and it is structurally unable to answer the questions a breach would.
Server-side aggregation
Counts and rates are computed inside TargetFlo. Row-level data is never serialized to the model.
Scoped tokens
Per-organization, per-role tokens reach the ops schema only. Revocable in admin at any time.
No PHI schema access
The MCP server has no connection path to client, guardian, insurance, or document tables.
Audit log
Every tool call — who, what, when, which parameters — sits in the same log as PHI access events.
Architecture note
How a question becomes an answer
Four steps, all inside TargetFlo's boundary.
- 01
Question arrives via MCP
Claude calls a TargetFlo MCP tool such as pipeline_stage_counts with parameters like location and date range. The tool list is fixed; there is no free-form query.
- 02
Token is checked
Each connection uses a scoped token issued per organization and per user role. Tokens can only reach the ops schema, never identity or PHI.
- 03
Aggregate is computed server-side
TargetFlo runs the count or rate against operational tables and returns a number, a table, or a short summary. Row-level data never leaves the server.
- 04
Call is logged
Tool name, parameters, requester, timestamp, and response shape are written to the audit log, alongside every other TargetFlo action.
Roadmap
Coming soonHelp shape the first release
The MCP server is designed and on the roadmap. The waitlist is where we learn which questions operations leaders actually ask, so the first set of tools matches a Monday morning meeting rather than a demo script.
- Early access when the read-only tool set ships
- Input on which aggregates and breakdowns come first
- Review of the token and audit design against your compliance requirements
- Works with Claude via the open Model Context Protocol
- pipeline_stage_counts(location?, from, to)
- stage_dwell_time(stage, location?)
- fax_queue_status(location?)
- task_summary(team?, status?)
- course_completion(course?, role?)
- eligibility_outcomes(from, to)
FAQ
AI & MCP Assistant FAQs
Related features
Join the MCP Assistant waitlist
Book a demo to see the live platform and tell us the three questions you most want answered on Monday morning. We will build the first tools around them.