disaster-safety-web-mcp

Inspiration

Traditionally, AI agents have struggled during disasters to provide guidance that simultaneously ensures safety and accommodates individual needs. By leveraging WebMCP, an AI agent can dynamically access accurate, real-time disaster information and mapping data while understanding a person’s specific circumstances.

Whether someone needs directions to a designated shelter, nearby high ground, or a destination compatible with mobility and transportation constraints, Untitled makes personalized evacuation guidance possible. We were inspired by the opportunity to move beyond one-size-fits-all evacuation orders and deliver practical, potentially life-saving support for each individual.

## What it does

Untitled is an AI-powered disaster safety assistant that turns live geographic and emergency data into personalized evacuation guidance.

Through WebMCP, the assistant can:

• Resolve natural-language place names while detecting ambiguous locations.
• Display flood forecasts and inundation zones on an interactive map.
• Find official evacuation shelters and classify them as clear, at risk, or unknown.
• Retrieve active alerts from authoritative agencies.
• Calculate walking or driving evacuation routes that avoid known flood hazards.
• Present map information as accessible text for screen readers and environments without WebGL.

Rather than simply naming the nearest shelter, Untitled considers whether that shelter and the route leading to it are actually safe.

## How we built it

We built Untitled as a TypeScript application running on Bun, with a MapLibre GL JS frontend and a Hono-based backend.

The browser registers disaster-response capabilities directly through WebMCP, allowing an AI agent to discover and invoke tools such as geocoding, flood forecasting, shelter search, official-alert retrieval, map control, and evacuation routing.

The backend integrates data from sources including the Japan Meteorological Agency, Japan’s Geospatial Information Authority, the US National Weather Service, MeteoAlarm, OpenStreetMap, Open-Meteo, and Copernicus services. Flood- aware routes are calculated using road-network data and dynamic hazard exclusions.

We also developed hydrological modeling components for rainfall runoff, terrain-based inundation, river capacity, and flood-risk estimation. Every result carries provenance so the system can distinguish live observations, planning scenarios, cached information, and test fixtures.

## Challenges we ran into

The greatest challenge was preventing the system from appearing more certain than its data justified. Disaster information can be delayed, incomplete, ambiguous, or temporarily unavailable, so we designed the application to report uncertainty explicitly instead of silently substituting a potentially misleading answer.

Combining datasets was also difficult because each provider uses different geographic resolutions, update schedules, risk scales, and coverage boundaries. Flood forecasting introduced additional complications: some climate datasets are rate-limited, while Copernicus retrievals can take hours and must be managed as resumable jobs rather than ordinary requests.

Mapping presented another subtle challenge. Data being present in application state does not necessarily mean it is visible on the rendered map. We therefore test the features that MapLibre actually draws after camera movement has completed.

Finally, accessibility required us to create a synchronized text representation of every visual map feature so the experience would remain useful to people using screen readers or devices without WebGL.

## Accomplishments that we're proud of

We are proud that Untitled provides an end-to-end workflow—from a natural-language request to hazard analysis, shelter evaluation, and a flood-aware evacuation route—through tools an AI agent can discover directly in the browser.

We are especially proud of the system’s safety boundaries. Official alerts are preserved as attributed source material, ambiguous locations are not silently guessed, unavailable forecasts are not presented as “no danger,” and developer-only controls are completely excluded from the normal user experience.

We also built a browser verification suite with 44 checks against the real application. It validates tool registration and invocation, error handling, accessibility, production-mode isolation, and what the map actually renders—not merely what the application intended to display.

For Japan’s experimental flood-risk forecast, we achieved 98.5% exact level agreement in evaluated reference areas while documenting the more difficult flood-extent limitations honestly. That distinction reflects a principle that guided the entire project: in disaster response, trustworthy uncertainty is more valuable than impressive but misleading confidence.

## What we learned

We learned that connecting an AI agent to live data is only the beginning. Useful disaster guidance requires provenance, freshness checks, spatial reasoning, uncertainty handling, and clear separation between observed hazards and planning scenarios.

We also learned that “nearest” rarely means “safest.” A shelter may be close but located inside a hazard zone, and a destination may be safe while the route to it crosses floodwater.

Most importantly, we learned that safety systems must test observable outcomes. A successful API response does not prove that a person can see the hazard, understand the warning, or follow the proposed route. Reliability has to be evaluated across the entire experience.

## What’s next for Untitled

Next, we want to expand coverage to additional countries and disaster types, including wildfires, earthquakes, landslides, tsunamis, and extreme heat.

We also plan to improve personalization for mobility needs, language preferences, family and pet evacuation, vehicle availability, medical requirements, and shelter accessibility. Offline and low-bandwidth support will make the assistant more resilient when communications infrastructure is damaged.

Built With

• ai