Amtrak GTFS + GTFS-RT MCP server

A public, read-only Model Context Protocol server that lets AI agents query both live and scheduled Amtrak transit data:

Realtime (GTFS-RT) — live predictions, train positions, and alerts, fetched on every call from the transitdocs feed.
Static (GTFS) — routes, stops, trips, schedules, and service calendars, loaded into this val's SQLite database from Amtrak's GTFS.zip.

Realtime tools

Tool	Description
`get_feed_header`	Feed metadata: version, timestamp, and counts of trip updates / vehicles / alerts.
`get_trip_updates`	Real-time arrival/departure predictions. Filters: `routeId`, `tripId`.
`get_vehicle_positions`	Live train positions (lat/long, bearing, speed, current stop). Filters: `routeId`, `tripId`, `vehicleId`.
`get_service_alerts`	Active service alerts / advisories. Filter: `routeId`.

Static GTFS tools (SQLite)

Tool	Description
`get_routes`	List routes. Filters: `routeId`, `query` (name search).
`get_stops`	List stations. Filters: `stopId`, `query` (name search), or `lat`/`lon`/`radiusKm` (nearby).
`get_trips`	List scheduled trips. Filters: `routeId`, `tripId`, `headsign`, `limit`.
`get_trip_schedule`	Full itinerary of a trip: stops in order with scheduled times. Requires `tripId`.
`get_stop_departures`	Scheduled departures/arrivals at a station. Requires `stopId`, optional `routeId`.
`get_calendar`	Service days/date range for a `tripId` or `serviceId`.
`get_agency`	Transit agency info. Filter: `agencyId`.

Realtime times are ISO-8601; static schedule times are HH:MM:SS; delays are in seconds.

Data freshness

All realtime tools (get_feed_header, get_trip_updates, get_vehicle_positions, get_service_alerts) share a single 60-second cache of the raw GTFS-RT feed, stored in SQLite. The upstream feed is fetched at most once per minute; every tool reads from the same cached copy.
Static GTFS is stored in SQLite and refreshed daily at 4:00 UTC by the refresh-gtfs.ts interval, which re-downloads the GTFS.zip and re-ingests via ingest-gtfs.ts. You can also run ingest-gtfs.ts manually at any time.

Connect an MCP client

Point your MCP client at the server's URL (streamable HTTP transport). For example, in a Claude Code / MCP config:

{
  "mcpServers": {
    "amtrak-gtfs": {
      "type": "http",
      "url": "https://<your-subdomain>.val.run"
    }
  }
}

You can also test it quickly in the Cloudflare MCP playground.

Example

Ask your agent: "Where is Amtrak train 287743 right now, and what are its upcoming stops?" It will call get_vehicle_positions and get_trip_updates (live) and get_trip_schedule (scheduled) to answer.

Notes

Read-only: realtime tools only fetch the upstream feed; static tools only read SQLite.
Built on the v2 TypeScript SDK (@modelcontextprotocol/server), compatible with MCP 2.0 (2026-07-28 spec) and the legacy spec.