MCP Healthcare Server

A Model Context Protocol (MCP) server for clinical workflows, built per the design in deep-research-report.md. The server exposes a governed set of tools (patient lookup, appointment booking, prescription, drug-interaction check, symptom triage, lab results, insurance eligibility, telehealth) and enforces role-based access control and audit logging on every call.

This implementation is a development scaffold: data is held in memory (seeded from src/data/*.json) and there is no real EHR/FHIR integration. It is a safe environment to exercise the MCP surface and the security enforcement layer.

Quick start

npm install
npm run dev     # tsc-watch-style restart on src changes
# or
npm run build && npm start

The server speaks MCP over stdio and is ready to attach to any MCP-aware client (Claude Desktop, MCP Inspector, custom agents).

Tool surface

Tool	Role required	Purpose
search_patient	any clinical	Look up a patient by MRN or name+DOB
get_patient_history	any clinical	Recent appointments, conditions, meds
find_available_slots	any clinical	Provider availability, filterable by specialty
book_appointment	clinician / admin	Schedule a new appointment
reschedule_appointment	clinician / admin	Move an existing appointment
get_active_medications	any clinical	Patient's current medication list
check_drug_interaction	clinician / pharmacist / admin	Pairwise interaction check
create_prescription	clinician / admin	Write a prescription, surfaces allergy + interaction warnings
analyze_symptoms	any clinical	Rule-based triage with urgency + specialty
suggest_specialist	any clinical	Recommend a specialty for a description
get_recent_lab_results	any clinical	Fetch lab data, filter by test name
analyze_lab_results	any clinical	Summarize, count abnormalities, flag criticals
check_eligibility	clinician / admin	Verify insurance coverage and copay
create_prior_authorization_request	clinician / admin	Submit a prior auth request
create_telehealth_session	clinician / admin	Create a video session, return URL
send_telehealth_link	clinician / admin	Queue session link for delivery (stub)

Every tool call must include a userId from src/data/users.json. The server resolves the user, checks role permissions, then runs the handler. Denials and runtime errors are recorded in the audit log.

Security model

• Audit log — one JSON line per call to logs/audit-<YYYY-MM-DD>.log. Captures timestamp, userId, tool, SHA-256 of the input (16 hex chars — avoids writing PHI in plaintext), outcome, and duration.
• RBAC — defined in src/auth/rbac.ts. Read tools accept any clinical role. Write tools require clinician or admin. Pharmacy tools also accept pharmacist. Unknown user → denial.
• Input validation — Zod schemas on every tool; the MCP SDK rejects malformed inputs before the handler runs.

File layout

src/
  index.ts                 # bootstrap (stdio transport)
  server.ts                # createServer(): McpServer + audit + exit hooks
  config.ts                # env loading (LOG_DIR, DATA_DIR)
  types.ts                 # domain types
  audit/logger.ts          # append-only audit log
  auth/rbac.ts             # user lookup + role checks
  store/memoryStore.ts     # in-memory store + loadSeed()
  tools/
    registry.ts            # central tool context + module list
    patient.ts             # search_patient, get_patient_history
    appointment.ts         # find_available_slots, book, reschedule
    prescription.ts        # create_prescription, check_drug_interaction, get_active_medications
    symptoms.ts            # analyze_symptoms, suggest_specialist
    lab.ts                 # get_recent_lab_results, analyze_lab_results
    eligibility.ts         # check_eligibility, create_prior_authorization_request
    telehealth.ts          # create_telehealth_session, send_telehealth_link
  data/                    # seed JSON (patients, providers, drugs, labs, etc.)
logs/                      # audit log output (gitignored)

Seed users (for testing)

userId	name	role
u-clin-1	Dr. Aisha Khan	clinician
u-clin-2	Dr. Marcus Lee	clinician
u-nurse-1	Nurse Priya Shah	nurse
u-recep-1	Jordan Park	receptionist
u-pharm-1	Rita Chen, PharmD	pharmacist
u-admin-1	Sam Whittaker	admin

Out of scope (TODO for production)

• Real FHIR / HL7 / PACS / DICOM adapters
• OAuth2 / SMART-on-FHIR launch
• Persistent database (in-memory only)
• LLM-backed symptom analysis (rule-based placeholder)
• HIPAA-grade audit retention, encryption at rest, key rotation
• Rate limiting, network segmentation

References

See deep-research-report.md for the full design rationale, architecture options, compliance checklist, and roadmap that informed this implementation.