Devpost Submission Fields

Inspiration

People already use AI for health questions every day. They upload lab reports to ChatGPT, ask about drug interactions, share photos of prescriptions. But every time, the AI starts with a blank slate. It doesn't know you're diabetic. It doesn't remember your medications. It doesn't know your cholesterol has been climbing for two years. Generic advice without context can be harmful. We asked: what if your AI actually knew you? "Nosce te ipsum" — know thyself. That's the gap we're filling.

What it does

Nosce is a patient-owned health context system. It gives AI assistants persistent memory about a patient's health using MCP, A2A, and FHIR.

Patients share medical documents in natural conversation. Nosce extracts clinical data (medications, lab values, diagnoses), saves it permanently as FHIR R4 resources with timestamps, and provides personalized preventive care recommendations and lifestyle plans based on the complete health history.

The system has three components: Nosce Core (MCP server with 9 health tools and full FHIR CRUD), Nosce Holistic (A2A lifestyle medicine specialist), and Nosce Intake (orchestrator agent that processes documents and coordinates care). Data persists across sessions and LLMs — connect from any AI assistant and your health record is there.

How we built it

Nosce Core is a Python MCP server built with FastMCP 3.x, exposed over Streamable HTTP. It declares support for Prompt Opinion's FHIR context extension and reads patient identity from SHARP headers. Google Gemini 2.0 Flash powers the clinical reasoning — generating preventive care gap analysis and personalized lifestyle plans from the patient's full FHIR record. Every record includes timestamps (effectiveDateTime, authoredOn, onsetDateTime). Two A2A agents are configured directly on Prompt Opinion: Nosce Intake as the patient-facing orchestrator, and Nosce Holistic as the lifestyle medicine specialist that gets consulted automatically when chronic conditions are detected. The server supports dual authentication — Prompt Opinion's automatic SHARP headers and personal API keys for direct MCP connections from any LLM client.

Challenges we ran into

FastMCP v3 changed its API significantly from v2 — constructors, app initialization, and transport setup all needed adaptation. The google-generativeai Python package was deprecated mid-build, requiring migration to the new google-genai SDK. Understanding how Prompt Opinion passes FHIR context via HTTP headers (not tool parameters) required implementing custom Starlette middleware with context variables. Declaring the FHIR extension in MCP capabilities required monkey-patching FastMCP's initialization options since there's no native API for custom extensions yet.

Accomplishments that we're proud of

Built a complete MCP + A2A solution that works end-to-end on Prompt Opinion. Successfully declared and activated the FHIR context extension — the platform recognizes our server as FHIR-aware and passes patient credentials automatically. Created a dual-auth system that works both on the platform AND as a standalone MCP server any LLM can connect to with a personal API key. The lifestyle recommendations are genuinely useful — they consider drug-food interactions, existing conditions, and give 3 actionable steps instead of overwhelming 10-point plans.

What we learned

MCP is genuinely powerful as a universal tool protocol — the same server works across platforms without modification. FHIR R4 maps naturally to what patients share in conversation. The combination of MCP (tools) + A2A (agent collaboration) creates something neither can do alone — tools persist the data, specialist agents reason over it. Prompt Opinion's SHARP context propagation solves the hardest problem in healthcare AI — identity and authorization — so builders can focus on the use case instead of plumbing.

What's next for Nosce

Offline-first with local AI. A rural patient with no consistent internet opens Nosce on their phone. Gemma runs locally, reads their prescription, saves to on-device storage. When WiFi returns, everything syncs to their cloud FHIR record. Their telemedicine doctor's AI agent on Prompt Opinion immediately sees the updated history. The sync pattern is simple — local data marked "pending" gets pushed to cloud FHIR when connectivity returns, then local copies are cleared. Beyond that: trend analysis ("Your A1C dropped from 7.2 to 6.8 over 6 months"), multi-language support via on-device Gemma, wearable integration from Apple Health and Fitbit, and one-tap provider sharing before doctor appointments.

Built With

• fastmcp
• ngrok
• prompt-opinion
• python