MCP Server

The Charter Boats MCP server lets Claude, ChatGPT, Cursor, and any MCP-compatible AI assistant search boats, locations, POIs, routes, trips, and content -- with rich HTML widgets for visual boat and destination cards.

Endpoint: https://charter.boats/mcp Install / quickstart page: open https://charter.boats/mcp in a browser. Transport: Streamable HTTP (JSON-RPC 2.0). Auth: None. The server does not pass request headers on to the API, so an API key sent to /mcp has no effect — MCP calls are neither attributed to an account nor affiliate-tagged. See also Rate limits. Spec: MCP Apps 2026-01-26.

Deployment

The server is a Cloudflare Worker routed at charter.boats/mcp. It proxies tool calls to the Charter Boats AI API (https://charter.boats/api/ai/*) and serves two HTML widgets as MCP App resources.

AI assistant → https://charter.boats/mcp (CF Worker) → charter.boats/api/ai/*

A single /mcp URL handles three things:

• GET with Accept: text/html → install / quickstart page (HTML)
• GET with Accept: application/json → server info JSON
• POST → MCP JSON-RPC

Tools

Ten read-only tools. Each returns the underlying API response as structuredContent (which the widgets render) plus a short text summary for the model. A search that finds nothing still returns the API's zero-results block — no_results (searched, why, try_next) and note — in structuredContent.

Search

search_locations answers three different questions, picked by which inputs you send:

• By name — q (min 2 characters): "Lefkada", "Dubrovnik".
• By popularity in an area — region (a sailing region as guests say it: Dalmatia, the Ionian, the Cyclades…), country, city, optionally type (marina, harbour, anchorage, bay, mooring, spot — the same words the results carry back, spot being an unclassified stop) and month (1-12, default the current month). Ranked by sort: popularity (default — observed vessel traffic for that month), boats (charter fleet size) or rating.
• By reach from a place — from (a location id or slug from an earlier result) plus within_nm (10, 20, 35, 55 or 100; default 35): the places boats were recorded sailing to from there, most-travelled first, with estimated distance and hours at 6 knots.

See Search Locations for the response of each mode.

Detail

Full input schemas are exposed via tools/list -- view them in MCP Inspector by connecting to https://charter.boats/mcp.

Prompt presets

The server registers three prompts surfaced as one-click starters in supporting clients:

• plan-sailing-trip -- destination + optional guests/dates/budget
• find-charter-boat -- location + optional type/guests/dates
• explore-destination -- single destination, surfaces marinas + POIs + routes

UI widgets

Two HTML widgets are served as MCP App resources:

The {hash} is a hash of that widget's own HTML, so the URI changes if and only if the widget does. ChatGPT caches MCP resources by URI with no expiry — a fixed URI means a redeployed widget is never re-read, however many times it ships — and OpenAI's guidance is to give the template a new URI when its HTML, JS or CSS changes. Read the current values from resources/list; never hard-code one. A read of a previously published URI (including the original unhashed ui://widget/boats.html) still returns the current widget rather than an error, so a host holding a stale URI degrades to old-metadata-but-correct-markup instead of a broken card.

Both widgets:

• Use the text/html;profile=mcp-app MIME type.
• Declare their image CSP twice over: _meta.ui.csp.resourceDomains (the MCP Apps standard, which Claude reads) and _meta["openai/widgetCSP"].resource_domains (the snake_case compatibility key, which is the only one ChatGPT reads — without it ChatGPT applies a locked-down default and every boat photo renders as a broken image). _meta.ui.prefersBorder / openai/widgetPrefersBorder and _meta.ui.resourceUri / openai/outputTemplate are paired the same way.
• List exact origins, never wildcards — ChatGPT rejects a *.example.com entry and one bad entry discards the whole list. The allowed set is charter.boats, media.charter.boats, our storage bucket's direct origin, and wsrv.nl. Only the boats widget shows images: boat photos are served from media.charter.boats (listed explicitly, since a CSP host source matches that host alone and charter.boats does not cover its subdomains), and a boat photo not yet copied to our storage resizes through wsrv. The locations widget shows no images, and search_pois has no widget — its image_url values are passed through as stored (some on our storage bucket, many at a third-party site's own address: a review site, a venue's website, a local guide), and nothing renders them inside a widget.
• Implement the full MCP Apps handshake: sandbox-resource-ready → ui/initialize (with protocolVersion, appInfo, appCapabilities) → ui/notifications/initialized → render on ui/notifications/tool-result.
• Report height via ui/notifications/size-changed, measured from body.scrollHeight / #app (the document element returns 0 inside Claude's sandbox).
• Route link clicks through ui/open-link since sandboxed iframes block target="_blank".
• Respect host theme: explicit theme: 'dark' | 'light' from host-context-changed overrides the prefers-color-scheme CSS fallback.

Install

For end users: open https://charter.boats/mcp in any browser -- the page has copy-paste install steps for Claude.ai, Claude Desktop, ChatGPT, Cursor, and MCP Inspector.

Quick reference

Claude.ai / Claude Desktop: Settings → Connectors → Add custom connector → https://charter.boats/mcp (Pro/Max/Team/Enterprise required).

ChatGPT: Settings → Apps → Create app → MCP Server → https://charter.boats/mcp. Widget rendering requires the ChatGPT Apps program; Developer mode shows tool calls as text only.

Cursor: add to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "charter-boats": { "url": "https://charter.boats/mcp" }
  }
}

MCP Inspector: npx @modelcontextprotocol/inspector → Streamable HTTP → https://charter.boats/mcp.

Rate limits

Tool calls are subject to the shared /api/ai/* rate limit. That limit is counted per calling client, and for MCP the caller the API sees is this server, not your assistant or its user: the server forwards no end-user identity (no IP, no key). MCP traffic is therefore limited collectively rather than per person, and there is no per-user figure to plan against. A limited call surfaces as a tool error (API 429); wait a little and retry.

Differences from the Custom GPT

Both integrations call the same /api/ai/* endpoints but expose them differently: