AI route optimization usually means a paid app with a sales call and a “per driver per month” price. What I built is smaller and weirder: a free connector that lets Claude plan multi-stop delivery routes for you, with a real solver behind it, no account and no signup. You type “plan tomorrow’s deliveries for my 3 vans” in a chat, paste your addresses, and get back routes, arrival times and Google Maps links. It’s called the Fleet Route Optimizer, and it has been live in the Claude directory for about a month.
This post does a few things at once (I tried to split it and it kept coming back together): how to use it, how I built it, what the usage log says after the first month, and what I’m changing because of it.
What it does, in plain words
An MCP server is a small web service that an AI assistant can call as a tool. MCP stands for Model Context Protocol, the open standard Anthropic published so assistants like Claude can talk to outside tools in a common way. The assistant is good at reading your messy message (“the bakery closes at 2, the van carries 40 boxes”); it is bad at the math of putting 30 stops in the best order for 4 vans. So it hands that part to my server, which runs the same engine as my free route optimizer website: Google OR-Tools with real road driving times.
What it handles:
- One depot, up to 100 stops and 25 vehicles per call
- Time windows per stop (“between 9 and 11”) and service time at each stop
- Up to three capacity limits per vehicle (boxes, weight, pallets, whatever you count)
- Street addresses (up to 50 looked up per call) or latitude/longitude
- Re-planning from the routes you already have, when one stop gets added or cancelled
What you get back: the stop order for each vehicle, the clock time it arrives at each stop, how much it’s carrying, and a Google Maps link per route you can send to the driver. If a stop can’t be served (the van is full, or the window is impossible) it says which one and why, instead of quietly dropping it. One honest note: these are the best routes the search found in the time allowed, not a mathematical proof that nothing better exists. Routing is one of those problems where proving it would take longer than the delivery day.
How to use AI route optimization in Claude
You need a Claude account. Then open the Fleet Route Optimizer page in the Claude directory and click connect. There’s nothing to log into on my side. After that, just ask. Some prompts that work well:
“Plan routes for 2 vans from my warehouse at 1500 E Douglas Ave, Wichita KS. Each van carries 40 boxes. Here are today’s stops with box counts and delivery windows: …”
“I have 12 addresses to visit tomorrow, one car, I start at 8 AM from home. Give me the best order and a Google Maps link.”
“Planificá rutas de reparto para 3 camionetas desde este depósito: …” (Spanish works too; there’s a Mexico City example built in.)
If you’re not sure what to give it, ask Claude to “show me the route optimizer example first”. It will pull a 30-stop Cleveland sample and you can see the shape of the input before you paste your own. A small tip from watching the log: give it the depot (where the vans start) up front. More on why below.
If you’d rather not use an AI at all, the same solver runs on the website with an Excel workbook, and that’s how most people still use it.
How I built it (with Claude doing most of the typing)
I’ve been writing optimization models for twenty-something years, mostly in Excel, and I’ve been moving them to the web one at a time. The routing engine itself is old news on this blog (the route optimization algorithm post explains how it actually searches). The MCP part was new to me, and I’ll say it plainly: I built it with Claude. I made the decisions, Claude wrote most of the code, and I tested it.
The decisions that mattered:
No login. Every routing tool I looked at wants an account or an API key. Mine doesn’t. Instead it’s rate-limited: one solve at a time on the server, a cap on calls per address, and size limits (100 stops, 50 address lookups, 25 vehicles). That’s what keeps a $7-a-month server alive.
Same server as the website. The MCP endpoint lives on the same Render service as routing.kindoflost.com, so one engine, one set of bugs.
Jobs, not one long request. A 100-stop plan can take 30 seconds or more, and AI assistants don’t like waiting that long on one call (nobody publishes the exact timeout, so this was a guess I had to test). So the tool waits up to about 25 seconds, and if the solve isn’t done it hands back a job number and Claude asks again a bit later.
Friendly defaults. People don’t type the way my Excel workbook wants. So if you don’t say how many vehicles, it assumes one; no capacity means unlimited; no time window means any time; times can be “8:30 AM”, “08:30” or minutes. Anything it doesn’t understand comes back as a warning instead of an error.
Then I published it to the official MCP Registry and submitted it to the Claude directory. It was approved the same morning as a Community connector. The whole thing, from “let’s do routing” to listed, took about two days, mostly because the scheduler version had already taught me the hard lessons a week earlier (I’ll write that one up separately).
How it’s going: the first month in the usage log
Every call to the server writes one row to a Google Sheet: when, how big, how long, and how it ended. No addresses, no names. Here is what the routing connector did between September 28 and October 24:
75 solved route plans sounds nice until you look closer (it always does). Eight of them are mine, from launch day on September 28, when I tested the connector before submitting it. When Claude calls my server the request comes from Anthropic’s computers, not from the person typing, so the log can’t tell me who anyone is. But I didn’t touch it after that day, so the other 67 plans were real people, spread over five different days. One session on October 3 ran about 40 one-vehicle plans in 20 minutes, which looks like Claude planning one driver at a time. That’s somebody actually using it (or somebody very bored).
The plans were small: most had 10 to 20 stops, the biggest had 99 stops and 7 vehicles. That fits what I expected, an AI chat is where you bring your Tuesday list, not your 500-stop fleet.
The interesting part is the failures, because each one is something I can fix:
43 empty calls. Every day, sometimes in bursts of three in the same second, something calls the tool with no data at all. These are directory health-checkers making sure the server answers. Harmless, but they were being logged as “bad input” next to real mistakes, which made the tool look broken when it wasn’t.
7 “busy”. The server runs one solve at a time. When Claude decided to plan several drivers at once and fired the calls in parallel, the extras got turned away. Claude retried, but the message it got didn’t really tell it what to do.
5 errors, and 4 of them were the same thing: a list of about 45 addresses where looking them all up took longer than the 45 seconds I allow. The funny part (well, funny now) is that the next try worked, because the addresses found the first time are saved. The error just never said so.
What I’m changing next
All of these come straight from the log above:
Empty calls now get a friendly answer that points to the example, and they’re logged as test calls, so I can count real mistakes separately. If you give stops but forget the depot, it will use the first stop as the depot and tell you it did, instead of refusing. The “busy” answer now tells the assistant to wait a few seconds and send one call at a time. The address timeout now says which addresses were already found and that trying again will be fast.
I’m also adding two things so I can learn what people actually want. The first is an optional “goal” field, one sentence in your own words of what you’re trying to do. Assistants tend to fill it in, and it tells me more than any count can. The second is a “check request” tool, so before trying, Claude can ask whether the server can handle something. Pickups and deliveries in the same route, several depots, or more than one trip per van are things it can’t do yet. Every time someone asks for one of those, I’ll know, and that list becomes my to-do list. Right now I’m guessing; the pickup and delivery route post was a guess too.
The connector is free and will stay free. If you plan deliveries for a small business, a food bank or just your own errands, try it and tell me what broke. I read every row of that log anyway.
Things that I use, like, and am affiliated with:
Mint Mobile offers great cell phone service for $15 flat, get $15 off using the link. Get discounted phones with service activation and no contract.
I never spend money before I check Mr Rebates or Rakuten to get cashbacks, rebates, discounts, coupons or cheaper gift cards.