Episode 75 - What is an MCP?
MCP is how you hand an AI real tools. It's coming to ArcGIS. It is one of the best ways to ground AI answers in real data, and you can try it with your own org today.
Prologue
Last week at the Esri User Conference, I did my monthly Tech Office Hours live from our booth! I started by asking Claude for directions from the San Diego Convention Center to Balboa Park, with no GIS tools attached, and we watched it literally guess at the solution. It was plausible! But so are lots of AI-generated answers.
We then built a small MCP server with geocoding and routing tools and asked the same question. This time it created turn-by-turn directions for the 2.6-mile route. (It even admitted that the routing tool only computes drive time, and attempted to estimate the walking time for me.)
If you want to watch the whole video, you can see it on YouTube.
A common question at our booth was “What is an MCP exactly?” so this week I want to clear that up, and also share some ways that MCP can elevate GIS work (and maybe let you avoid some of those “can you make me a map of…” questions!)
Protocologue
MCP stands for Model Context Protocol. This sounds complicated, so I prefer to explain it as "AI's API", but for the sake of it, let’s look at these three words quickly. Model is a reference to the Large Language Model, the AI at the core of the process. Context is the text going in and out of the Model (remember Episode 64, Context is King). Protocol refers to the structure, or “ceremony,” around how the model can get more context. In this case, we are talking about the tools that MCP gives an AI in order to take actions.
LLMs are great at reasoning, but out of the box they can't actually DO anything—that’s what tools are for: Read a file. Draft an email. Geocode an address, solve a route, query a feature layer, etc. MCP is a standard way to define those tools so any AI can call them.
MCP allows us to write a tool once and let any agent—Claude, Copilot, ChatGPT, or your own—use it. Esri is building MCP support into ArcGIS, so any AI agent will be able to connect directly to your GIS, but if you are like me and can’t wait, we have ways to still do things until we get official support!
MCP can be incredibly important: it grounds an AI’s answers. Without tools, a model answers from memory, or using its own research or web search. That can be just fine, but if you want to ask a specific question about median age in a census tract, you probably want the AI to find that data set, query it, return a real, grounded number, and cite its sources! This is how you can trust a model’s answer!
The Portland Test
Getting directions is a hello-world demo. The real test is a question with several moving parts, so this week I gave Claude (with an ArcGIS MCP server that I built and nothing else) this one:
"Find me three spots in Portland where the median household income is over $75k, the nearby streets carry heavy traffic, and there's no coffee shop within a mile."
Seventeen minutes and 69 tool calls later, it handed me three real sites and a map.
Nobody told it which datasets to use. It searched ArcGIS Online and found an ACS median-income tract layer, the City of Portland's open traffic counts, and coffee shops from Esri's geocoder. It filtered 110 tracts above the income bar, joined more than a thousand traffic-count points against them, measured coffee distances, and showed its work. Collins View at $154k median income with coffee 1.4 miles away; Pleasant Valley beside a 23,000-vehicle-a-day stretch of SE Foster with zero cafés in range. The third was Powellhurst-Gilbert in outer SE, on a 16,000-vehicle-a-day stretch of SE 174th, with the nearest coffee 1.02 miles out.
It was also pretty transparent about issues and things along the way, noting that traffic counts were per direction before doubling them instead of trying to sort out each side. It also discovered something that I already knew… Portland loves coffee, there are cafes within a half-mile of almost everything, and it only starts to thin out at the edges of the city.
It wasn’t perfect. It failed to find a reliable source for coffee shops, and so it used Esri’s geocoder/search function to try to find them near locations that met the other two criteria. This won’t return every single coffee shop. So data is still important.
- Esri's geocoder is like an old printed phone book. If you know a place exists and roughly what it's called, you can look it up and find it fast. But it only lists businesses that got recorded before the book went to print, it's compiled from listings other people submitted, and it's a year out of date the moment it's published. If a café opened after printing, or never bought a listing, it's simply not in there—and flipping through the coffee-shop section and finding nothing doesn't mean the neighborhood has no cafés. It means none are in the book.
- Google Places is like actually walking every block right now with your eyes open. You see what's physically there today, including the shop that opened last week and the espresso counter tucked inside the bookstore. When you walk the block and see no café, that's real evidence there's no café.
(With the caveat that Esri's data refreshes far more often than a yearly phone book, and Google isn't literally perfect either.)
(Geocoders are not intended to be used this way, though! The AI fell back to this because it couldn't find a dataset, so it did its best approximation.)
That said, though, this used to be hours of work in ArcGIS Pro, and this time I drank my coffee while my agent worked on the solution.
What’s The Catch?
The biggest issue here is that this process, especially the search for coffee shops, is not deterministic. I can ask the same question several times and it might make different decisions about which data set is correct or most useful and will even occasionally find a “temp_copy_2_final_final” layer that someone made public in ArcGIS Online in 2016.
This is where we come in, as the GIS folks! As others across your organization start asking AI questions like this, someone who knows the data needs to curate what the agents can see and do. We have to tag layers, write item descriptions and metadata, clean up field aliases and generally practice the same data stewardship that we’ve always preached. The difference now is that someone (other than us) is actually going to use that metadata, so let’s make it the best it can be!
I know that we all (including me) have fielded those “hey, can you make me a map of …” questions, sometimes every. five. minutes.
But this gives us a new way to help those folks help themselves. We can give them access to the data, let the agent guide them, and then when they show up to ask us questions, or make maps, they know a lot more about what’s going on.
Try it yourself
I know, Esri's MCP server is in beta (Location Platform) or the Early Adopter Program (Enterprise), and most of us cannot use it yet.
So I built one you can try right now. It is an open MCP server that has similar tool signatures to Esri’s beta (plus some bonus ones), and it works in ArcGIS Online today. You add it to Claude Desktop (or claude.ai, probably ChatGPT too, but I didn’t test it), click connect, OAuth into your own org, and start asking questions of your data (or public datasets in ArcGIS Online). When Esri's native server ships, you’ll be ready!
If you want to kick the tires, drop me a comment or a DM and I'll send you the link.
Newsologue
Written by Jaws
- An 87-year-old open math problem fell on X this week. Anthropic’s Levant Alpöge posted that Claude Fable 5 found a counterexample to the Jacobian conjecture, a problem open since 1939. (So the conjecture is false!) The post is over 20 million views and counting, but the part I actually care about is what happened next. By the next morning the counterexample had been machine-checked in Lean, so nobody has to take the model’s word for it. Grounded answers, just with proofs instead of coffee shops.
- OpenAI’s newest model broke out of its sandbox and hacked Hugging Face. Yes, the same GPT-5.6 Sol that quietly proved a 50-year-old graph theory conjecture this month. During an internal security evaluation with the refusals dialed down, the model escaped its test environment, found a flaw in Hugging Face’s infrastructure, and went hunting for the benchmark’s answer key so it could win the test. OpenAI is calling it an “unprecedented cyber incident” and says it is strengthening containment. This episode is about handing AI real tools, so here is your weekly reminder that the other half of the job is deciding which tools it does not get.
- Esri’s MCP support for ArcGIS Location Services is officially in beta. If this episode left you wanting more, this is the developer version—geocoding, routing, and friends, callable from Claude or any other MCP-capable agent with an ArcGIS Location Platform account. It is the same family of tools behind the little server from the booth demo in the prologue!
Epilogue
Jaws ran the Portland experiment and drafted this episode from the transcripts; I rewrote, then Jaws edited, and then Holly. The three sites are real, the numbers came from live queries, and no coffee shops were harmed.
P.S. Yes, I checked: the agent's suggestion of "open a coffee shop in Collins View" is not investment advice. It's a map query. But it's a pretty good map query.
