Independent / MCP & ChatGPT
Swiss Transit Explorer
An MCP server and ChatGPT app that turns Swiss transit data into practical travel decisions.
- Contribution
- Design and development
Independent MCP project
Source available · hosted demo offline / 2026
The challenge
A timetable can list connections, but it cannot explain which journey is the best fit for a traveller. This project set out to turn Swiss transport data into useful recommendations, accounting for tight transfers, delays, walking exposure and arrival buffers. It was also our first complete MCP server and ChatGPT app: the tools had to accept natural-language constraints reliably, return compact results and render usable widgets, while running on an emerging stack whose production behaviour differed from local development.
The work
Starting on a train ride to Italy, we built an MCP server with mcp-use and Swiss public transport open data, then developed it into a route-planning app inside ChatGPT. We designed focused tools for connections, station boards, disruptions and delays, added route ranking and transfer-risk scoring, and built interactive route and departures widgets. Optional weather and GTFS line mapping made the results more useful. Beyond the first working demo, we refined tool selection, arrival-time parameters and token usage, simplified the interface, and resolved production failures involving widget assets, build ordering and ESM imports.
- Composable MCP tools for connections, departures, arrivals, disruptions and route delays
- Route ranking with transfer risk, delay context, walking exposure and arrive-by buffers
- Optional Open-Meteo weather and GTFS mapping from raw identifiers to readable transit lines
- Interactive route explorer and departures-board widgets with concise risk indicators
- Prompt and parameter refinement for arrival times, transfer limits and detail levels
- Production debugging for deployment, widget asset paths, build sequencing and explicit ESM imports
The result
A complete MCP server and ChatGPT app with ranked connections, explanations and interactive widgets, supported by a detailed account of the implementation and deployment lessons. The hosted demo is currently unavailable; the recorded demo, source and local setup preserve the work.