MCP server: the route optimizer inside your AI assistant
https://routing.kindoflost.com/mcpFree, no signup, no API key. Streamable HTTP transport.
Why use it
A language model can order ten stops “by eye” and hand back a route that looks reasonable, crosses town twice, and reaches a customer after they closed. With this server the assistant turns what you tell it into the input, and the routes are computed by the same engine that runs in the web version: Google OR-Tools over real road driving times (OpenStreetMap), with several vehicles, time windows, service times and up to three capacity limits per vehicle. Any stop that cannot be served is named instead of silently dropped. It is a heuristic search: you get the best routes found in the time allowed, not a proof of optimality.
How to connect
Claude: at https://claude.ai/customize/connectors add a custom connector with the URL above. No OAuth needed.
Claude Code: claude mcp add --transport http route-optimizer https://routing.kindoflost.com/mcp
Other assistants and clients: anything that accepts a remote MCP server over HTTP; use the same URL.
What to ask
For example: “Two vans leave our warehouse at 1200 Main St, Columbus, OH at 8am. Plan routes for these 18 deliveries; the pharmacy on High St only takes deliveries 9–11, and each drop takes 10 minutes.” Or: “Here are yesterday's routes plus three new stops — re-optimize them.”
Tools
list_route_examples— the built-in problems (30 deliveries around Cleveland, Ohio; 24 around Mexico City).get_route_example— the full input of one example, to edit into your own.optimize_routes— plans the routes and returns, per vehicle, the stop order with arrival times and load, the stops that could not be served, totals and Google Maps links.get_route_result— collects a long solve that did not finish inside the first call.
Limits
Up to 100 stops, 50 addresses to look up and 25 vehicles per call; one depot, deliveries only, one trip per vehicle. A run takes 10 to 60 seconds, longer when many addresses must be looked up (lat/lon is faster). The server solves one problem at a time, so at a busy moment the assistant may be asked to retry in a minute. For bigger problems use the web version.
Privacy
Addresses without coordinates are looked up with the US Census geocoder and OpenStreetMap, and coordinates are sent to the public OSRM server for driving times. Addresses that are found are kept on the server with their coordinates so they are not looked up again. Everything else is used only to compute the routes and is discarded within 15 minutes. Each run writes one anonymous log row, with no addresses or names. Details in the privacy policy.
Support: mario@kindoflost.com.