Project

Campus Map Chat

UC San Diego · GIS · AI

AI agents are getting better at handling busywork, like sorting email, scheduling meetings, or making purchases. But at times, the location context is overlooked, and a resulting AI-planned schedule might expect you to teleport from one side of campus to the other in five minutes (a walk that would typically take 20). I built an MCP server that lets AI agents like Claude query UC San Diego's geospatial data, including buildings, points of interest, and walking routes.

Meanwhile, finding the nearest lunch spot or a walk that skips the stairs on a campus map usually means a lot of pinching, zooming, and guessing. An assistant can take on that legwork and base its recommendations on real campus data.

I prototyped this while working at UCSD, where the MCP server called the university's ArcGIS APIs directly for current map data. The public demo below runs on a saved snapshot of those layers, so please double check anything important before relying on it.

Try it

Try it in the web app

https://map-chat.alessioyu.xyz/mcp

Some prompts to try

1. Coffee on the way "I'm in Geisel Library and need to get to class in Podemos. I'd like to get coffee on the way. Can you suggest a route that includes a coffee stop?"

2. Where do I park? "I'm driving in to visit Geisel Library. Which parking structures are closest that have visitor spaces, and how long is the walk from the nearest one?"

3. No stairs, please "I'm at the Price Center and can't use stairs. Is there a step-free walk to Geisel Library?"

What it can do

Search 752 buildings and more than 3,700 points of interest, from coffee carts to gender inclusive restrooms
Pull up details on any place
Plan walking routes over campus paths, sidewalks, crosswalks, and stairs, with a step-free option
Chain routes across several stops

How it works

Data
- Features: a snapshot of UCSD's public campus map POI and Buildings layers
- Routing: a walking network built from OpenStreetMap paths
Web app agent
- The agent model is Claude Haiku
- Answers are grounded in facts pulled from the tools instead of memory, to reduce made-up places
Rate limiting
- Web app messages have per visitor hourly limits and a bot check on every prompt
- The rate limit also applies to the MCP endpoint
- If you hit a limit while trying the app or MCP out, please try again later

Where this could go

This demo is aimed at students and visitors, but the same architecture could support more operational work. The idea is to expose a few well-defined tools over data an organization already keeps in ArcGIS, and let an assistant combine them to support faster, better-informed decisions.

For example, since UCSD's maintenance requests already live in ArcGIS, facilities staff could ask "Which open requests are within a five minute walk of this building?" and then "Group the plumbing ones into one walking loop." That pairs a request query with the same routing tool used here. A staff version would need sign in, authorization levels (read-only, update permissions), and a human approving anything that changes data.

For visitors, an AI agent could pull live transit feeds and suggest when to leave for the next shuttle so you don't have to keep checking.