SwitchBack

Inspiration

I am usually short on time for planning my next hike or trail run. I always leave it until the very end. It is hard to Google "ChatGPT" and get routes from AllTrails or Wikiloc. They are unverified, and more than once I have been on closed hikes or dangerous segments. Also, I think we need to be able to co-create with agents.

Planning a hike is not really a question with a single text answer. It is a shared, spatial decision: where to start, whether a route genuinely closes as a loop, how far it is, what the ascent looks like, and what the current planning context is.

I built Switchback around a simple idea: ask for a loop, then see the ground truth. Rather than having an agent recommend a walk somewhere in a random place, a person and an agent work on the same Collserola–Vallvidrera map and can inspect the same route, evidence, and limitations.

What it does

Switchback is a WebMCP-native trail planner for short, car- or public transport-accessible walks around Collserola in Barcelona, Spain. It is the closest national park to the town.

A person can ask for a family loop, a specific distance, or an evening option in natural language. The agent discovers the WebMCP tools, selects from graph-verified circuit options, validates a route before changing the map, and renders the result directly in the shared interface.

The result stays inspectable:

A closed circuit is drawn on the map.
The route card shows measured distance and sampled ascent.
Forecast, daylight, and official Park notices appear as clearly labeled planning context.
Every agent action is visible in a shared worklog and trail ledger.
The agent can prepare a briefing or GPX trace, but the person remains in control: they review/copy the briefing and manually click Download GPX.

How I built it

Switchback is a static TypeScript app hosted on Netlify. It registers fourteen page-bound WebMCP tools through the browser’s model context, with strict JSON schemas and the same planner functions used by the manual UI.

Routing happens locally in the browser over a 66-tile directed Collserola TrailPack assembled from OpenStreetMap data and the Park’s published public-use network. The planner only offers routes that the available graph can close without a long retraced leg. I have used the official PDF and the park's website.

I added route-aware forecast/daylight context via Open-Meteo and official Park-alert context through a same-origin adapter. Original Catalan notices remain visible; English is clearly labeled as machine translation. I used the AWS Translate service and Amazon Location services for the translations and maps.

Challenges I ran into

The hardest problem was avoiding an impressive-looking but untrustworthy route suggestion. A distance target alone does not prove that a loop closes, and a forecast does not prove that a hike is safe.

That led to a few important product constraints:

Routes are graph-validated before they render.
Short, Medium, and Long are distance profiles, not difficulty labels.
Difficulty is only described conservatively when the available ascent and terrain data support it.
Forecasts and Park notices are planning signals, never a safety clearance.
The agent cannot auto-download GPX files, send messages, or access external accounts.

What I learned

WebMCP is most compelling when the agent is not treated as a chat layer bolted onto an app Sometimes, it was hard for ChatGPT to use the tools because the agent's skill was not working correctly. I even fixed my local installation of ChatGPT with Codex.