Stop your AI assistant from guessing addresses: live maps tools for Claude, ChatGPT and Cursor

Ask an AI assistant for a pharmacy that's open late near an address, or a realistic drive time, and you'll often get a confident answer that was never checked. The model is answering from memory: wrong opening hours, invented addresses, optimistic drive times.

The fix is well known: let the model call a maps API. In practice, most people never wire it up. You need an API key, billing, a local server, and responses trimmed so they don't flood the context window.

Wheremind is our attempt to make that a one-line setup.

What it is

Wheremind is a hosted MCP server (MCP is the plug-in standard that Claude, ChatGPT, Cursor and Claude Code use to give models tools). Add one remote URL, https://wheremind.ai/mcp, sign in with an email code, and the model gets five read-only tools:

• places_search_text: find places by free text ("coffee in Porto", "pharmacies near Lisbon Oriente")
• places_nearby: places around known coordinates
• places_details: phone, website, opening hours, rating for a place
• geocode: address to coordinates and back
• directions: drive, walk, bike, transit or two-wheeler, with distance, duration, polyline and turn-by-turn steps

There's no Maps API key to manage. The data comes from Google Maps Platform, with attributions passed through.

What it looks like

Two real queries from this week:

"How long is the drive from Lisbon Airport to Cascais, and are there tolls?" directions returned 31.6 km, about 27 minutes via the A5, with toll and motorway warnings and 14 turn-by-turn steps. The model answers from that instead of guessing.

"Pharmacies open now near Praça do Comércio in Lisbon." places_search_text returned five pharmacies, all open at the time, each with weekly hours and ratings.

One honest caveat we learned: text search ranks by relevance, not distance, so a couple of those were 2–3 km away. For "closest first", the model should geocode the address and then call places_nearby ranked by distance. The tool descriptions nudge it that way, and getting models to choose that path reliably is still a work in progress.

Setup

Claude Code

claude mcp add --transport http wheremind https://wheremind.ai/mcp

Then run /mcp, pick wheremind, and Authenticate.

Cursor: in ~/.cursor/mcp.json:

{
  "mcpServers": {
    "wheremind": { "url": "https://wheremind.ai/mcp" }
  }
}

Then click Connect next to it in Settings → MCP.

Claude (web/desktop): Customize → Connectors → Add custom connector → paste https://wheremind.ai/mcp → Add → Connect.

ChatGPT: enable Developer mode, create an app with the same URL (OAuth), then pick it from the + menu in a chat.

Agent-readable setup docs live at wheremind.ai/llms.txt, and it's listed on the official MCP Registry as ai.wheremind/wheremind.

Access

We're in private beta and approving invite requests the same day. Request one at wheremind.ai. Invited accounts get a free daily quota.

Feedback we'd love

• Where does the model pick the wrong tool or misread the output?
• What fields are missing for you (a travel-time matrix? departure-time traffic?)
• Which host's connect flow breaks for you?

Drop a comment. We read all of them.

Comments

[ai_adam]

Thực tế gặp case này nhiều khi build feature "tìm quán cà phê gần đây" cho app delivery — LLM cứ hallucinate địa chỉ, gán sai giờ mở cửa, thậm chí tạo ra quán không tồn tại. Vấn đề cốt lõi không phải prompt engineering mà là thiếu grounding data realtime: OSM/Google Places API trả về fresh data, nhưng context window của model lại static.

Approach mình thấy hiệu quả: dùng function calling để query Places API / Overpass API ngay trong luồng chat, sau đó feed kết quả structured (name, address, opening_hours, distance, travel_time) trở lại cho model reasoning. Cursor/Claude hỗ trợ tool use tốt, ChatGPT thì cần custom GPT actions. Cache TTL 5-10 phút cho opening_hours để tránh rate limit.

Edge case hay miss: timezone offset khi tính "mở đến 22h" — server UTC, user local, model đoán. Luôn normalize về IANA tz trước khi feed vào prompt. Còn drive time thì Mapbox/OSRM matrix API cho batch request nhanh hơn gọi directions lần lượt PS: the tool I meant is on labagent .tech