Inspiration

ICU handover is a high-stakes coordination problem: vital signs change, teams need a shared view of risk, and every action needs a clear owner. We wanted to explore a more trustworthy role for an AI agent in that workflow.

Wardround is built around one principle: the agent can prepare the next action, but a nurse remains the person who commits it. Instead of treating AI as an autonomous decision-maker, we designed it as a visible, accountable co-pilot.

Wardround uses synthetic demo data only. It is not medical software and must not be used for clinical care.

What it does

Wardround is a simulated ICU shift-handoff and triage co-pilot built with React and WebMCP.

The live ward board updates simulated patient vitals every four seconds, calculates risk from configurable thresholds, and ranks patients by urgency. An agent in ChatGPT's in-app browser can use scoped WebMCP tools to:

• identify patients by clinical risk;
• explain the top concern and recent trend;
• propose a patient flag, note, alert acknowledgement, or manual triage order;
• draft a shared, risk-sorted handoff summary.

Clinical workflow changes are never applied by the agent directly. Write tools create a pending proposal with visible clinical provenance. A nurse can then approve or reject it in Wardround, and every proposal and decision is recorded in the audit trail.

The shared handoff is the one deliberately narrow exception: an agent can draft it, while the nurse can edit and explicitly save it, turning it into a co-authored shift artifact.

How we built it

Wardround is a client-side React application built with Vite, JavaScript, CSS, and Zustand.

• A synthetic vital-sign feed updates the ward state every four seconds.
• A threshold-driven risk engine scores heart rate, SpO2, respiratory rate, temperature, and blood pressure, then produces a severity level and a visible top concern.
• WebMCP tools are registered through document.modelContext and read live Zustand state only when invoked.
• Two read tools return tightly scoped patient-risk information.
• Four proposal tools can request clinical workflow changes, but can only add to the approval queue.
• A separate handoff tool updates only the shared handoff draft and cannot alter patient, triage, or approval state.
• The interface includes live risk cards, approval controls, provenance, a persistent audit log, a shared handoff surface, and keyboard-accessible dashboard views.

Wardround is deployed on Cloudflare Pages and can be opened inside ChatGPT's in-app browser, where the agent discovers and calls the WebMCP tool surface.

Challenges we ran into

The hardest challenge was not making an agent call tools. It was making the safety model structural.

We had to ensure that a write-capable tool could never bypass human review. The solution was to make every clinical write tool a thin translator from agent intent into a proposal object. Only the nurse approval path can apply a proposal to patient workflow state.

Other challenges included:

• keeping tool results scoped and schema-tight;
• ensuring tools always read the current live ward state rather than stale data;
• preserving a nurse-approved manual triage order while the live feed continues to recalculate risk;
• making alert acknowledgements specific to the active concern so an old acknowledgement does not hide a new alert;
• creating a responsive dashboard that remains usable in desktop, split-pane, tablet, and mobile contexts;
• designing deterministic demo mode so the core workflow is reliable to show in a short presentation.

Accomplishments that we're proud of

We are proud that Wardround is more than a dashboard with AI language around it. The agent, interface, and audit model are connected through a real WebMCP tool surface.

In particular, we are proud of:

• seven intentional WebMCP tools with clear boundaries;
• human-gated clinical proposals rather than autonomous workflow changes;
• visible provenance that explains why the agent made a suggestion;
• a real approval and rejection loop with patient-state changes only after approval;
• a shared handoff that preserves agent authorship, nurse edits, and audit history;
• a risk board that stays live while the human decision process remains explicit;
• a polished, responsive command-center interface designed for fast clinical scanning.

What we learned

We learned that trustworthy agent design is as much about product architecture as it is about prompting.

A good tool description is not enough. The strongest safeguard is designing the system so that the tool cannot take an unsafe shortcut in the first place. Separating “propose” from “apply” made the safety story easy to demonstrate, easier to audit, and easier for a user to understand.

We also learned that responsive UX matters for agentic products. A judge might experience Wardround in a full browser, a narrow split pane beside ChatGPT, or a mobile-sized viewport. The product has to preserve clarity and control in all three.

What's next for Wardround

Next, we would add authenticated roles, real clinical-system integrations, persistent audit storage, role-based escalation policies, and clinician-configurable thresholds.

Before any real-world use, Wardround would require clinical governance, privacy and security review, validation against real workflows, and extensive safety testing. The current project is intentionally a synthetic prototype that demonstrates the interaction model: AI can help teams see and prepare; humans remain accountable for action.

Built With

• chatgpt
• cloudflare
• css
• html
• javascript
• react
• vite
• webmcp
• zustand