Hormuz Watch — MCP Server

Tactical maritime situation dashboard for the Strait of Hormuz — two interactive Leaflet map widgets + three JSON tools, delivered as a single MCP endpoint. Fuses IMF PortWatch chokepoint statistics with live AIS vessel positions into one picture you can query in natural language from Claude, ChatGPT, or any MCP‑Apps client.

Ask it: "Is Hormuz open right now?" · "What ships are at Fujairah?" · "Where does Hormuz rank against Suez and Malacca today?"

The two widgets

show-hormuz-situation — strait overview

Everything you need to answer "what's happening in Hormuz?" in a single pannable map: IMF disruption polygon, 65 Gulf ports color‑coded by country and sized by annual traffic, 150 live AIS vessels inside the strait bbox, a 30‑day transit sparkline, telemetry chips (transits vs. historical baseline, Δ‑below‑normal, live vessel count), and a RED/ORANGE/GREEN alert banner driven by IMF PortWatch.

show-port-live-map — single-port AIS recon

Focused radius scan around any Gulf port. Crosshair marker, scan ring, every vessel inside the radius (heading, type, stale flag), ordered ledger sorted by range, and a composition breakdown by vessel class.

What the server exposes

All 5 tools visible in the mcp-use Inspector:

Tool	Kind	Description
show-hormuz-situation	widget	Primary dashboard — full strait overview with live AIS, ports, disruptions, trend.
show-port-live-map	widget	Radius-scan AIS map around one port (e.g. port362 Fujairah, port744 Jebel Ali, port306 Dubai).
get-hormuz-disruptions	JSON	Active/recent IMF trade‑disruption events affecting Hormuz with alert level, dates, description.
get-port-traffic	JSON	Daily ship‑call history for one Gulf port with per‑vessel‑type breakdown.
compare-chokepoints	JSON	Latest daily transits across all 28 global chokepoints, rankable by total or by vessel type.

How it works

Data sources

IMF PortWatch — live ArcGIS FeatureServer layers. No API key, no rate limits in normal use, ~3–7 day reporting lag, cached 1h in process.

Layer	Used for
Daily_Chokepoints_Data	Daily transit counts for 28 global chokepoints (Hormuz = chokepoint6).
PortWatch_chokepoints_database	Historical baseline (annual totals by vessel type).
portwatch_disruptions_database	Active/historical trade‑disruption events with alert level + polygon geometry.
PortWatch_ports_database	65 Gulf port metadata.
Daily_Ports_Data	Per‑port daily ship‑call counts.

VesselFinder /api/pub/mp2 — live AIS vessel positions (MMSI, name, lat/lon, heading, type). Undocumented binary endpoint used by VesselFinder's own front‑end, reverse‑engineered in lib/vesselfinder-client.ts. 30‑second cache. ToS‑gray: may be rate‑limited, IP‑blocked, or break without notice. For production deployment, swap to a commercial AIS API (Datalastic, Spire, VesselFinder paid tier). The rest of the dashboard keeps working if AIS fails — fetchLiveShips has a catch fallback that returns an empty vessel list.

Terrestrial AIS receivers have effectively zero coverage of the Persian Gulf, so free public AIS streams deliver no messages from Hormuz. The satellite-assisted VesselFinder feed was the only free source that actually returned vessels here.

Payload shaping

The strait widget needs a lot of data, and the MCP tools/call response ships it all as structuredContent. To keep that responsive the server trims before returning: disruption polygons are decimated to ≤50 points, live ships are capped at 150 nearest to the chokepoint, history is reduced to {date, total}, and unused fields on latest, chokepoint, and disruption are dropped. See decimatePolygon / nearestShips in index.ts.

Architecture

index.ts                       # MCP server + 5 tools (2 widget + 3 JSON)
lib/
├── portwatch-client.ts        # ArcGIS fetchers for 5 PortWatch layers + 1h cache
└── vesselfinder-client.ts     # Binary AIS decoder for /api/pub/mp2 + 30s cache
resources/
├── hormuz-map/                # Strait overview widget
│   ├── widget.tsx             # SSR‑safe entrypoint (McpUseProvider + lazy-loaded map)
│   ├── map-inner.tsx          # Leaflet map + alert banner + telemetry + sparkline
│   └── types.ts               # Zod schemas for widget props
├── port-live-map/             # Single‑port AIS widget
│   ├── widget.tsx
│   ├── map-inner.tsx
│   └── types.ts
├── shared/theme.ts            # Dark tactical theme + ship/country/alert color tables
└── styles.css                 # Tailwind-generated widget styles

Rendering: OpenStreetMap/CARTO tiles via Leaflet + react-leaflet. Fully interactive — pan, zoom, click popups, tooltips, layer toggles. No map API key. No billing.

Widget contract: each tool declares widget: { name: "hormuz-map" | "port-live-map", invoking, invoked } at registration. mcp-use auto-wires openai/outputTemplate → ui://widget/<name>.html and registers the matching resource under resources/<name>/widget.tsx. The tool returns widget({ props, output }); the props land in useWidget().props inside the iframe.

Built with mcp-use

This project is a pure mcp-use server. What that gets you:

• server.tool(...) — single-call registration with Zod schema, async callback, and a widget config block that links to a React widget folder under resources/.
• widget({ props, output }) helper — returns a CallToolResult with structuredContent populated and automatically picked up by the widget iframe via useWidget().
• mcp-use dev — starts the MCP server and a Vite dev server for the widgets, with HMR, type generation (.mcp-use/tool-registry.d.ts), and an Inspector at /inspector.
• mcp-use build + mcp-use deploy — one‑shot production build + deploy to mcp-use Cloud.
• MCP Apps + OpenAI Apps SDK dual protocol — the same widget renders in Claude (MCP Apps), ChatGPT (Apps SDK), and the Inspector with no per‑client code.

Nothing in this repo calls the raw MCP SDK — it's all import { MCPServer, widget, object, text, error } from "mcp-use/server" and import { useWidget, McpUseProvider } from "mcp-use/react".

Install & run

npm install
npm run dev

Open http://localhost:3000/inspector, pick show-hormuz-situation, click Execute. The interactive map renders with the live RED disruption polygon, all 65 Gulf ports, and every AIS contact inside the strait bbox.

Connect to Claude / ChatGPT

Start a tunnel from the Inspector (Start Tunnel button) or deploy, then add the MCP URL to your client:

• Claude Desktop — Settings → Developer → Edit Config → add the server under mcpServers.
• ChatGPT Apps — add the HTTPS endpoint as a custom connector.
• Inspector — already connected, just hit Add to Client.

Deploy

npm run deploy

Pushes a production build (Vite bundle of both widgets + server) to mcp-use Cloud and returns a public HTTPS URL.

Layout conventions

• resources/<widget-name>/widget.tsx must export default a React component; the directory name is what the tool's widget.name config points at.
• widgetMetadata.exposeAsTool: false on both widgets — the tool is declared explicitly in index.ts, not auto‑generated from the widget.
• Zod schemas in each widget's types.ts are the single source of truth for widget props; the server-side widget({ props: … }) object must match.

License

MIT