I Gave Claude AI Full Control of QGIS: What Happened
I copied a shapefile path into a chat box and asked Claude to build a map. A few seconds later, QGIS opened a new project, loaded a Seattle bus network, zoomed to the extent, and reported the feature count back to me. I never touched the QGIS interface.
That is what AI in QGIS looks like right now. Not a button inside the software. A separate AI app that reaches into QGIS and runs the same tools you would run by hand. I wanted to know how far it goes, where it falls apart, and whether it actually saves you anything. So I wired up the QGIS MCP server, connected Claude, and ran a real workflow on real data.
What the QGIS MCP plugin actually is
MCP stands for Model Context Protocol. It is an integration that lets a language model like Claude talk directly to another application. In this case, the QGIS MCP plugin opens a two-way connection between the Claude desktop app and QGIS.
Once it is connected, Claude can create projects, manipulate layers, run processing algorithms, and execute code inside your QGIS environment. The project was built by Juan Sashawaki, and it covers project manipulation, layer manipulation, and processing operations.
The important part is what is happening underneath. There is no magic. If you read the plugin code, every action runs through the QGIS Python package. The LLM reads your instruction, picks a tool, and that tool fires well structured Python to do the work. Claude is the interpreter. Python is the engine. That distinction matters when things break, because the errors you hit are Python and QGIS errors, not AI errors.
Setting up Claude QGIS: the steps that trip people up
Getting this running takes a handful of steps, and two of them cause most of the trouble.
Install UV and clone the repo
UV is a Python package manager, similar to pip. Once UV is installed, you download the repo. I use the GitHub CLI to clone it, but you can grab it any way you like.
Move the plugin into your QGIS plugins folder
Inside the repo there is a QGIS MCP plugin folder. You have to copy that folder into the location where QGIS stores plugins on your machine. On a Windows machine you can do this from the command line. On other systems you point to your QGIS application support folder where the rest of your plugins live. You can also zip the folder and install it through the manual plugin upload, which lands it in the same place.
Then close QGIS, reopen it, go to Plugins, then Manage and Install Plugins, and find the MCP entry. It starts unchecked. Check it to install it.
Connect the Claude desktop app
Install the Claude desktop app, then open Claude, go to Settings, then Developer, and edit the config. You paste in the configuration with one change that catches people: you have to set the absolute path to the parent repo folder, the one that holds the source QGIS MCP folder. That path tells the MCP server where the Python file lives that runs the whole thing.
After both steps, restart QGIS so everything connects. Restarting QGIS is the single most common fix when the connection does not show up. When it works, clicking the tools search in Claude shows QGIS as active, and you can choose which tools to give it access to.
Running a real AI GIS automation workflow
I loaded two Seattle datasets to test it: a transportation network with bus and transit lines, and a DEM for elevation.
Loading and centering a layer
I copied the shapefile path and asked Claude to create a new map, center on the data, and add it. Telling it where the data lives is not optional. The path is the instruction. My first run failed because I had not actually started the MCP server, so I asked it to set up QGIS MCP and start the server, then ran again. This time it created a project called Bus Network Project, added the layer, reported the feature count, and zoomed to the extent.
Styling and filtering by an attribute
Next I asked it to add an OSM base map, style the map by the speed limit column, come up with its own color scheme, and filter to records where bus flow is greater than zero. Naming the exact column, speed limit, works far better than describing it loosely.
Claude added the base map, then started generating its own classification code, defining color ranges between zero and 25 miles per hour. The result was reasonable but not perfect. The bus flow values in the data looked off, which may have been a data issue rather than an AI issue, but I got usable custom symbology that I did not write by hand.
Where it broke
The DEM step is where things got rough. I asked it to add the DEM raster and create a hillshade. It found the algorithm, generated the hillshade, and added it. Then, while trying to fix the blend mode and layer ordering, it removed the layer I wanted, could not relocate the DEM, and started throwing basic errors as it tried to re-add and restyle. This is the honest part of the test. The further you push into multi-step styling, the more the model adds extra steps and stumbles.
Should you use AI in QGIS?
Everything Claude did here, you can do by pointing and clicking. The plugin does not unlock new capability. It runs the same QGIS tools through code. So the value is not in raw power. It is in how you use it. I see three modes.
The first is running work in the background while you do something else. Honestly, this is the least valuable. If you already know the task, you do not need an AI to do it.
The second is advanced styling and exploration. The custom symbology it generated was genuinely useful. You can point at a map style you want, ask it to build that out, and skip the tutorial grind.
The third, and the most valuable, is learning. For the things you do not know how to do in QGIS, the AI works like a tutorial teacher built around your exact task. It shows you the steps and builds them in your map at the same time.
That last point is the real takeaway. If you only use AI to automate work you already know, you get a small convenience. Point it at the things you do not know yet, and it becomes a way to learn faster.
Frequently asked questions
What is the QGIS MCP plugin?
It is a plugin plus an MCP server that connects the Claude desktop app to QGIS. Once installed, Claude can create projects, add and style layers, and run processing tools. Every action executes through the QGIS Python package, so the AI is interpreting your request and choosing tools, not bypassing QGIS.
Do I need the paid version of Claude for QGIS?
The free version works, but you may only get a few commands before you hit limits. I used the paid version because it gives more credits and lets the model run more steps in a session. The constraint is your Claude account, not the plugin.
Does Claude QGIS replace knowing GIS?
No. It does not remove you from the process. You still have to understand your data, name the right columns, and check the output. It is strongest as a way to learn and to handle styling you would otherwise look up, not as a replacement for GIS skill.
Why does the AI keep hitting errors in QGIS?
Because it runs real QGIS and Python operations, it hits real QGIS and Python errors. Multi-step styling, blend modes, and locating raster paths are where it stumbles most. Restarting QGIS resolves most connection issues, and naming files and columns precisely cuts down on the rest.
If you want the full process for setting up these prompts, giving the model a system prompt, and using AI across QGIS, GeoPandas, and more, that is what I teach inside the Spatial Lab. Join here.
