02 — MCP integration

The app speaks MCP to upstream agencies, not bespoke HTTP clients. Each agency or data source gets one Model Context Protocol server; the agents in 03-agentic-architecture call them by name. New agency partners join the system by shipping a server, not by negotiating bilateral integrations.

MCP server inventory

Server Status Backed by Feeds vertical
vectorsurv-mcp ✅ shipped api.vectorsurv.org (OpenAPI v1.0.44) VBD
nws-heatrisk-mcp ⏳ planned api.weather.gov HeatRisk + alerts Heat
mag-hrn-mcp ⏳ planned hrn.azmag.gov (cooling-center registry) Heat
adhs-mcp ⏳ planned ADHS open-data (heat mortality, arbovirus reports) Both
great-az-tick-check-mcp ⏳ planned UA Cooperative Extension submission tracker VBD
whispers-mcp ⏳ planned USGS WHISPers wildlife mortality events VBD
inaturalist-mcp ⏳ planned iNaturalist API (vectors + wildlife observations) VBD
211-az-mcp ⏳ planned 211 Arizona resource directory Heat
knowledge-graph-mcp ⏳ planned This repo’s DuckLake graph (read + write) Both
outbreaks-near-me-mcp 🔭 future Federation with Boston Children’s Outbreaks Near Me symptom-cluster platform (successor to Flu Near You). OBNM is the canonical human-symptom-side participatory-surveillance platform; this MCP would let the AZ agents pull symptom-cluster context for a user’s ZIP and contribute Sentinel reports back upstream. Gated on a partnership with the HealthMap team. Both
medsis-mcp 🔭 future / blocked AZDHS MEDSIS communicable-disease surveillance — no public read API (PHI behind connect.azdhs.gov); only feasible as an outbound HL7/ELR reporting connector. See the MEDSIS assessment below. Both

Parameter-by-parameter source map

Where each Minimum-Dataset parameter is allowed to be populated from upstream (vs. always coming from the user):

General class

Param Source MCP
Geographical coordinates Device GPS — (client)
Postal code Device GPS reverse-geocoded → county/tribe lookup knowledge-graph-mcp (regions)
Date of report App timestamp —
Household member ID User-entered —

Human class — symptoms

Always user-supplied; the Enrichment Agent adds derived context (e.g. SNOMED CT code, severity score) from schema/deep/standards.sql. No MCP populates symptoms.

Auxiliary class — digital biomarker

Param Source MCP
Wearable skin-temp, HRV, sweat rate HealthKit / Health Connect on device — (client) — eventually wearable-mcp proxy
Photo Camera —
Lab confirmation User-entered + (later) FHIR HL7 pull future fhir-mcp

Environmental class — pulled live

This is where MCPs do the heavy lifting:

Param MCP Endpoint
Ambient temperature nws-heatrisk-mcp api.weather.gov gridpoint forecast
Humidity nws-heatrisk-mcp same
Heat index nws-heatrisk-mcp computed client-side from T + RH
NWS HeatRisk level nws-heatrisk-mcp api.weather.gov/products/heatrisk
Active heat warning nws-heatrisk-mcp api.weather.gov/alerts/active
Vector density (mosquito) vectorsurv-mcp vectorsurv_calculate_abundance
Vector positivity (WNV pools) vectorsurv-mcp vectorsurv_calculate_infection_rate
Tick observations nearby inaturalist-mcp obs?taxon=Ixodida&lat=…&radius=…
Wildlife mortality events nearby whispers-mcp events?bbox=…
Standing water / recent rainfall future nws-precip-mcp precip grid

Livestock / Wildlife class — both user and MCP

Param MCP
Species ID (from photo) knowledge-graph-mcp (taxonomy lookup) + Claude Vision
Number of sick/dead animals User-entered
Date / location of incident Device GPS + user
Cross-reference to current outbreak knowledge-graph-mcp + whispers-mcp + vectorsurv-mcp

Cooling-center awareness (Heat-Q1) — multi-MCP join

Heat-Q1’s “where can I cool off?” answer is a join across three MCPs and the knowledge graph:

user_location  →  mag-hrn-mcp.search_centers(lat,lon,radius)
              →  211-az-mcp.utility_assistance_nearby(zip)
              →  knowledge-graph-mcp.match_focus_area("cooling_centers")
              →  knowledge-graph-mcp.population_priority(user.demographic)
              →  ranked list of {center, distance, services, transport_available}

The Notification Agent renders the result with the highest-priority match (per the Q4 vulnerable-populations data) first.

VBD triage (Wildlife-Q4) — multi-MCP join

A community VBD report goes through:

user_report   →  vectorsurv-mcp.list_test_targets()
                 (look up matching pathogen by symptom + vector)
              →  vectorsurv-mcp.get_pools(near user.county, last 30 days)
                 (any positive pools nearby?)
              →  whispers-mcp.events(bbox=user.bbox, last 90 days)
                 (any wildlife die-offs nearby?)
              →  knowledge-graph-mcp.outbreak_check(pathogen, county)
                 (active outbreak record?)
              →  knowledge-graph-mcp.resource_lookup(pathogen,
                                                    user.tribe?)
                 (Great AZ Tick Check, AZGFD, ADHS, ITCA-TEC)
              →  Triage Agent decision: self-care | see clinician |
                 report to AZGFD | mail tick to Walker lab

MCP server spec template (for each new server)

To keep new MCP servers consistent with vectorsurv-mcp:

mcp/<name>/
  ├── README.md                 # what it does, install, env vars
  ├── pyproject.toml            # mcp[cli] + httpx + pydantic
  ├── .env.example              # credentials + PATH_* overrides
  ├── examples/
  │   └── claude_desktop_config.json
  ├── openapi/                  # versioned upstream-spec snapshots
  │   └── snapshot-<version>.json
  ├── src/<pkg>/
  │   ├── __init__.py
  │   ├── __main__.py           # stdio + streamable-http entry
  │   ├── client.py             # paths in env-overridable PATHS dict
  │   └── server.py             # FastMCP tools + resources
  └── tests/
      └── test_*.py             # synthetic-data tests, no live creds

Every server also drops a schema seed under schema/deep/ documenting its tools as graph nodes (mcp_server, mcp_tool, api, operatedBy, wraps, exposedBy, informs) so the knowledge graph itself can be queried for “what MCP tools answer Heat Q1?”.

Auth + data-sovereignty notes

MEDSIS (AZDHS disease surveillance) — integration assessment

MEDSIS is the Medical/Disease Surveillance Information System, the Arizona Department of Health Services’ official communicable-disease surveillance system of record (clinicians, labs, and county health departments report reportable conditions into it).

Can we build a medsis-mcp?

Not as a read-only data source — today this is blocked. Unlike vectorsurv-mcp (open OpenAPI) or adhs-mcp (ADHS open data), MEDSIS exposes no public query API; all access is credentialed PHI behind connect.azdhs.gov. A read MCP would require an AZDHS data-sharing agreement plus MEDSIS service credentials, and would be returning PHI — incompatible with this project’s anonymous, ZIP-coarse privacy contract.

The only feasible direction is outbound reporting — a connector that formats Sentinel signals as HL7 v2.5.1 ELR2PH / case-report messages for submission to ADHS, making the Sentinel a reporting source rather than a data consumer. This is gated on:

  1. a formal AZDHS partnership + ELR sender enrollment + HL7 validation; and
  2. a privacy reconciliation — MEDSIS case/lab reports are identified PHI, whereas Sentinel reports are anonymous and coarsened. Anonymous, ZIP-coarse observations are not reportable case reports, so any feed would be aggregate / syndromic-style signal counts agreed with ADHS, never row-level PII. The Validation Agent’s suppression rules and the Auth + data-sovereignty notes apply.

Recommendation: keep adhs-mcp (ADHS open data) as the near-term ADHS integration. Treat medsis-mcp as a future, partnership-gated outbound reporting connector, tracked as 🔭 future / blocked in the inventory — not a Phase-1 data source. If a partnership lands, build it from the spec template with an HL7 v2.5.1 ELR2PH message builder + a MEDSIS_* credential set, and add a sunset / MOU-renewal clause per GOVERNANCE.md (as with tribal-data servers).

Sources: AZDHS MEDSIS User Guide; AZ ELR Implementation Guide v1.3 (HL7 v2.5.1 ELR2PH); AZ Syndromic Surveillance Implementation Guide; AZDHS MEDSIS / Infectious Disease Services.