80% of pregnancy-related deaths are preventable. 40% of new mothers never return for postpartum care. MamaGuard closes the coordination gap with three specialist AI agents, 16 shared FHIR tools, approval-gated write planning, and the clinician always in control.
Built for the Agents Assemble hackathon. Dual-track submission: A2A Agent + MCP Server (Superpower), published on the Prompt Opinion Marketplace.
📄 MamaGuard_judge_packet.pdf — judge-facing narrative, scorecard, and demo walkthrough.
git clone <repo-url> && cd mamaguard
make setup # install deps (~15s)
make test # 1191 unit tests (<3s)
make tier1 # 59 deterministic benchmark cases, 100% pass rate
# To run the agent locally (needs GOOGLE_API_KEY in mamaguard/.env):
cp mamaguard/.env.example mamaguard/.env # then add your key
make serve # A2A agent on http://localhost:8001
make smoke # hit the running agent with 3 clinical queriesWhat you'll see: three specialist agents (maternal, pediatric, SDOH) coordinating through an orchestrator, producing structured 5T assessments with FHIR evidence citations, clinician review blocks, and confidence scoring -- all from real FHIR patient data.
MamaGuard is a multi-agent AI care coordination system that analyzes FHIR patient records to support maternal and pediatric health. Three specialist agents coordinate through a single A2A interface:
| Agent | Role | Key Tools |
|---|---|---|
| Maternal Risk Monitor | BP trends, glucose control, pregnancy history, postpartum complications | get_bp_trend, get_glucose_trend, get_pregnancy_history, get_maternal_risk_profile, write_risk_assessment |
| Pediatric Transition | Immunization gaps (CDC schedule), developmental milestones (AAP Bright Futures), care gaps | get_immunization_gaps, get_developmental_screening_status, get_care_gaps |
| SDOH & Outreach | Insurance coverage gaps, language barriers, food/housing insecurity, community resource lookup, actionable referrals | get_sdoh_screening, find_sdoh_resources, write_care_plan, create_communication_request |
Every tool response includes a Liaison Agent pattern clinician_review object -- AI recommends, clinician decides. When review is required, the agent pauses (INPUT_REQUIRED state) until the clinician acts.
Prompt Opinion Platform
/ \
A2A JSON-RPC MCP (SSE/stdio)
+ FHIR context + SHARP params
| |
v v
+-----------------------+ +------------------+
| A2A Agent | | MCP Server |
| Orchestrator | | (Superpower) |
| +-- Maternal Agent | +--------+---------+
| +-- Pediatric Agent | |
| +-- SDOH Agent | |
| Shared FHIR Tools <--- shared ---+
+-----------+-----------+
| HTTPS + Bearer Token
v
FHIR R4 Server
Data flow: Clinician launches MamaGuard from Prompt Opinion > selects patient > PO sends FHIR context (server URL, bearer token, patient ID) via A2A metadata > orchestrator routes to specialist agents > agents query FHIR server using shared tool layer > results include structured risk assessments with evidence citations > clinician reviews and decides.
| Metric | Current Baseline | Target with MamaGuard |
|---|---|---|
| Postpartum follow-up completion | 60% attend | 85%+ via proactive gap detection |
| Postpartum hypertensive crisis detection | Median 5 days | Same-day automated BP trend flagging |
| Childhood immunization adherence | 70.4% on time | 90%+ via automated gap detection |
| Clinician chart-review time | 15-20 min manual | <2 min AI-synthesized risk summary |
| SDOH screening completion | <25% of eligible | 80%+ via automated Z-code analysis |
Cost impact: Preventable maternal morbidity costs $32.3B/year in the US. Each avoided severe maternal morbidity event saves ~$115K in acute care costs.
-
Cross-resource synthesis -- A rule engine flags BP >140/90. MamaGuard looks at 6 pregnancies with 5 losses, uncontrolled HTN across 8 years, concurrent diabetes, and a single postpartum visit -- then explains why this combination is dangerous.
-
Contextual medication safety -- HCTZ is acceptable for HTN, but in a postpartum patient with diabetes who may be breastfeeding, the recommendation shifts to labetalol. This requires patient-specific clinical context.
-
SDOH-clinical integration -- Connecting a Medicaid gap + French language + food insecurity + postpartum BP crisis into a unified outreach plan with culturally appropriate referrals.
-
Natural language care plans -- Converting structured FHIR data into prioritized, evidence-cited care summaries using the 5T framework (Talk, Template, Table, Task, Transaction).
See benchmarks/clinical_reasoning/ for quantified AI Factor evidence: rule-engine baseline vs. MamaGuard synthesis across low/moderate/severe cases.
- Python 3.11+
- Google AI Studio API key (free)
# Clone and install
git clone <repo-url> && cd mamaguard
pip install -r mamaguard/requirements.txt
# Configure
cp mamaguard/.env.example mamaguard/.env
# Edit .env: set GOOGLE_API_KEY and MAMAGUARD_API_KEY
# Run the A2A agent (from repo root)
uvicorn mamaguard.app:a2a_app --host 0.0.0.0 --port 8001
# Agent card: http://localhost:8001/.well-known/agent-card.json
# Or run the MCP server (stdio -- for Claude Desktop, Cursor)
python -m mamaguard.mcp_server.server
# Or run with SSE transport (for remote clients)
MCP_TRANSPORT=sse MCP_PORT=8080 python -m mamaguard.mcp_server.serverdocker build -t mamaguard -f mamaguard/Dockerfile mamaguard/
docker run -p 8001:8001 \
-e GOOGLE_API_KEY=your-key \
-e MAMAGUARD_API_KEY=your-agent-key \
mamaguardFull local stack with a HAPI FHIR R4 server for end-to-end testing:
# Configure
cp mamaguard/.env.example mamaguard/.env
# Edit .env: set GOOGLE_API_KEY at minimum
# Start MamaGuard + HAPI FHIR
cd mamaguard && docker compose up -d
# Load 11 benchmark patients (127 FHIR resources) into HAPI
cd .. && python3 scripts/load_bundles.py --verify
# Verify
curl http://localhost:8001/.well-known/agent-card.json # Agent card
curl http://localhost:8090/fhir/Patient?_count=5 # FHIR patients
# Run e2e benchmarks against local stack
HAPI_FHIR_URL=http://localhost:8090/fhir python3 -m benchmarks.runner --e2e# Full pipeline: pre-deploy checks + deploy + verify
./scripts/deploy.sh
# Or dry run
./scripts/deploy.sh --dry-run# Unit tests (1191 tests)
python3 -m pytest mamaguard/tests/ -v
# Tier-1 deterministic benchmarks (59 cases)
python3 -m benchmarks.runner
# Pre-deploy check (unit tests + benchmarks, exit 1 if score < 95%)
./scripts/pre_deploy_check.sh
# With HAPI FHIR smoke tests
./scripts/pre_deploy_check.sh --hapiTest coverage:
- 1191 unit tests across 32 test modules
- 59/59 Tier-1 benchmark cases at 100.0%
- Agent routing, tool invocation, error paths, FHIR base utilities, FHIR writeback, SMART tickets, middleware, golden-file contract tests, LLM-as-judge care plan checkers, FHIR hook, logging utilities, mother-child handoff, MCP protocol integration, SDOH resource classification, A2A app factory and agent card endpoint, benchmark harness infrastructure, post-processing safety filter
Patient/bench-maria-001 -- a 50-year-old Black, French-speaking woman with:
- 6 pregnancies (5 losses), emergency delivery
- Uncontrolled HTN (BP consistently >140/90, spike to 170/98 postpartum)
- DM2 with HbA1c trending upward
- Only ONE postnatal visit, no insurance on record
MamaGuard identifies: URGENT risk (Stage 2 HTN + recurrent pregnancy loss + no coverage), generates a prioritized care plan with FHIR evidence citations, flags clinician review, and creates outreach resources including community referrals.
mamaguard/
+-- orchestrator/agent.py # Routes to sub-agents via AgentTool
+-- maternal_agent/agent.py # Maternal Risk Monitor
+-- pediatric_agent/agent.py # Pediatric Transition Agent
+-- sdoh_agent/agent.py # SDOH & Outreach Agent
+-- mcp_server/ # Standalone MCP server (Superpower track)
| +-- server.py # FastMCP, 19 tools, stdio + SSE
| +-- context.py # FhirContext adapter (SHARP -> state)
+-- shared/
| +-- tools/ # 21 shared tool functions (single source of truth)
| | +-- fhir_base.py # Patient summary, active medications, current plan, linked newborn
| | +-- maternal.py # BP trend, glucose, pregnancy history
| | +-- pediatric.py # Immunizations, developmental screening
| | +-- sdoh.py # SDOH screening, resource lookup
| | +-- writeback.py # Direct RiskAssessment, CommunicationRequest, CarePlan writes
| | +-- plan_mode.py # Approval-gated plan/commit write flow
| +-- app_factory.py # A2A app + AgentCard creation
| +-- middleware.py # API key auth + FHIR metadata bridging
| +-- fhir_hook.py # before_model_callback -- credential extraction
| +-- smart_tickets.py # SMART Permission Tickets (Mandel spec)
| +-- sdoh_resources.py # Offline SDOH resource map
+-- marketplace/ # Marketplace submission configs (both tracks)
+-- tests/ # 1191 unit tests
+-- app.py # A2A entry point
+-- Dockerfile # Cloud Run deployment
+-- docker-compose.yml # Local dev stack (MamaGuard + HAPI FHIR)
+-- requirements.txt
benchmarks/
+-- runner.py # Tier-1/2/3 benchmark harness
+-- clinical_reasoning/ # AI Factor evidence (rule engine vs. synthesis)
+-- llm_eval/ # LLM-as-judge care plan checkers
+-- e2e/ # End-to-end cases + FHIR bundles (11 patients)
+-- medagent/ # MedAgentBench (42 cases)
scripts/
+-- deploy.sh # Cloud Run deploy pipeline
+-- pre_deploy_check.sh # Pre-deploy verification
+-- load_bundles.py # Load FHIR bundles into HAPI (11 patients)
| Component | Choice | Why |
|---|---|---|
| Agent framework | Google ADK + A2A SDK | Reference implementation for A2A protocol |
| LLMs | Gemini 3 Flash, gpt-5.4 (production via PO routing); Nemotron 3 Super, Gemma 4 (benchmarks); DeepSeek 3.2 (LLM-as-judge) | Free/sponsor-credit tiers; benches run locally on DGX Spark |
| MCP server | FastMCP | Superpower track, dual submission |
| Healthcare data | FHIR R4 | Industry standard; SMART + HAPI servers |
| FHIR context | SHARP extension | EHR session credentials in A2A/MCP metadata |
| Permission model | SMART Permission Tickets | Josh Mandel's March 2026 spec |
| Deployment | Google Cloud Run | Single container, HTTPS, auto-scaling |
| SDOH resources | findhelp.org / 211 gateway | External directory with curated offline fallback |
- Liaison Agent pattern -- AI recommends, clinician decides. No autonomous clinical action.
- Credentials never in prompt -- FHIR tokens extracted via
before_model_callback, stored in session state, never visible to the LLM. - API key authentication -- X-API-Key header enforcement on all endpoints except agent card discovery.
- SMART Permission Tickets -- Scope-limited tool authorization (feature-flagged). Each FHIR-facing tool is mapped to required SMART v2 scopes.
- FHIR AuditEvent -- When enabled (
MAMAGUARD_AUDIT_EVENTS=true), every tool invocation emits a FHIR R4 AuditEvent (HIPAA compliance trail). - Minimum Necessary -- Each tool queries only the FHIR resources it needs.
- Synthetic data only -- SMART R4 Synthea records. No real PHI.
Python 3.11 | Google ADK | A2A SDK | Gemini 3 Flash | gpt-5.4 | Nemotron 3 Super | Gemma 4 | DeepSeek 3.2 | FHIR R4 | FastMCP | SMART/HAPI | Google Cloud Run | Docker | httpx | uvicorn | PyJWT
ARCHITECTURE.md-- Full architecture, agent definitions, security model, implementation planmamaguard/marketplace/-- Marketplace setup guides, system prompts, demo scriptmamaguard/mcp_server/README.md-- MCP server docs, tool reference, Docker
Licensed under the Apache License, Version 2.0. Copyright 2026 Michal Kurc.