What it solves
- Ranks options by transfer risk, delay signals, and outdoor exposure.
- Provides station boards, disruption checks, and route-specific delay context.
- Wraps everything in ChatGPT widgets instead of forcing long text responses.
Case Study
Swiss Transit Explorer
A decision-ready Swiss public transit MCP server and ChatGPT app that turns raw timetable data into ranked route options, live station boards, disruption checks, and weather-aware guidance.
Stack
Built with mcp-use, ChatGPT widgets, Swiss public transport open data, Open-Meteo, and optional GTFS line mapping for human-readable train and bus labels.
The hosted demo endpoint is currently unavailable. Run it locally →
Related write-up
The build log covers ideation, tool design, prompt routing, token issues, and the deployment problems that surfaced when moving from inspector to production.
Demo
Chat-first route planning.
Widgets
Tools that stay readable inside ChatGPT.
Route Explorer
Risk-ranked route options with transfer details, leg breakdowns, and clearer decision cues than a standard timetable UI.
ChatGPT App Surface
The final app combines tools, widgets, and structured prompts so ChatGPT can call the right actions without drowning the user in raw API output.
Inspector Workflow
The inspector was the fastest way to validate tool parameters, widget resources, and prompt behavior before deployment.
Lessons
Most of the hard work was not the routing logic.
Token hygiene
Weather became optional, prompts were consolidated, and the UI was simplified so the app could stay useful without wasting context.
Deployment friction
Deploying from source, fixing asset paths, and tightening build order were all necessary before the production app behaved like the local inspector build.
Outcome
The final build shipped ranked routes, route and station diagnostics, optional weather context, and a deployable MCP endpoint for ChatGPT integration.
Next
Need an MCP app or agent workflow?
I can help design the tool surface, widget UX, and deployment path so the app feels like a product instead of a protocol demo.