Add ground_location_tool for LLM grounding use cases

[mattpodwysocki]

Overview

Add a ground_location_tool that fans out to multiple Mapbox APIs and returns a structured, cited grounding response for a given coordinate + query. This enables LLMs to answer location questions from live Mapbox data rather than training data.

Motivation

Google recently launched Grounding with Google Maps for Vertex AI. Mapbox can offer a differentiated version using navigation, routing, and terrain data that Google doesn't expose.

Proposed Tool: ground_location_tool

Input:

{
  "latitude": 37.7749,
  "longitude": -122.4194,
  "query": "family-friendly restaurants within walking distance",
  "language": "en"
}

Internally composes:

• reverse_geocode — place/neighborhood context
• category_search — nearby POIs matching query
• isochrone — walkable/driveable area
• Optional: static_map_image — visual context

Output:

{
  "place": "SoMa, San Francisco, CA",
  "nearby_pois": [...],
  "travel_times": { "walk": "5min", "drive": "12min" },
  "citations": ["mapbox-search", "mapbox-directions"]
}

Related

• Companion skill tracking issue: mapbox/mapbox-agent-skills
• The skill will serve as the spec/prototype before this tool is built