GIS is finally going native.
Most GIS work is clicking through menus to produce a layer you could describe in one sentence. With NativeGIS you describe it, and your AI assistant builds it with real geoprocessing and shows its work.
Built for the assistant you already talk to.
NativeGIS is a local MCP server for LiDAR, raster and vector data. MCP is the standard way an AI assistant like Claude calls tools. GIS software was built for people with a mouse. NativeGIS is built for your assistant, so the work starts from a sentence instead of a toolbar.
You know the layer you need. Building it takes a day.
The problem
You need 40 sampling plots across a stand, stratified by canopy height, none on a slope over 30 percent, none within 50 m of a road. Or harvest zones that keep a stream buffer. Or three treatment scenarios for a client on Friday.
In a desktop GIS that is a chain of buffers, clips, joins and random-point tools, each with its own dialog. You build it once, tweak a constraint, and build it again.
With NativeGIS
Ask for the layer. The assistant picks the tools, runs them on your data and returns a real layer. Change the constraint and it reruns.
Because every step is recorded, the second run takes seconds, and the first is something you can read and check.
Ask for sampling plots, management zones, buffers, corridors and scenarios. You describe the intent. The geometry is computed, not drawn.
The data you have is gappy, noisy and not quite clean.
The problem
Real data has holes. A DEM with voids under water or in steep shadow. A point cloud with birds, noise and a ground class that is too sparse. Polygons that overlap by a sliver or leave gaps between neighbours. Stand records with no height, no species, no age.
You can fix each by hand, one dataset at a time, and the cleanup is the part nobody budgets for.
With NativeGIS
Ask for the enhancement and see what changed. Fill DEM voids. Clean point clouds. Fix overlaps and gaps in polygons. Add to each stand polygon the height, canopy cover and tree count measured from LiDAR.
Each step reports what it touched, so cleaned data stays traceable to the original.
Part of this already works. NativeGIS builds a ground surface from a LiDAR cloud, makes a terrain model from it and normalises heights above it. Where the delivered ground is too sparse it re-classifies the ground itself, and it keeps the vendor's ground when that is better. On a 56 ha test window in HJ Andrews, the 2008 terrain model built this way is within 0.23 m RMSE of a terrain model from a denser 2021 survey.
You can't tell how good a layer is until you compare it with something.
The problem
A canopy height model looks right on screen. A tree layer looks plausible. A DEM from a new survey looks like the old one. "Looks right" is not an accuracy figure, and a client will ask for one.
You need a second survey, a set of check points or a field plot to compare with. Setting that up is a project in itself, so most layers ship without it.
With NativeGIS
Ask for the comparison. Elevation change between surveys. Positional accuracy against surveyed check points. Trees found against trees measured in the field: omission, commission and height error.
You get numbers with the method stated, not a screenshot.
This is how we check ourselves. In the HJ Andrews test window we compared our results against independent software on the same data:
- Terrain surfaces built from the same ground points matched an independent triangulation to 0.000 m.
- The coordinate transform matched the PROJ library to under 1 cm in every test, across 18 projection pairs including the Albers projection US fire maps use (worst case 9.3e-9 m).
- The canopy height model matched an independent implementation to 1.01 m RMSE (r = 0.998) once both stood on the same ground.
- Tree tops: 4.1 percent of trees missed and 4.1 percent extra, against that same independent run.
We also say where it is weak. Trees under a taller crown are hard to see from above. In synthetic forests with known truth we found all the dominant trees but only 7 to 22 percent of the suppressed ones. Compare against field plots before you rely on stem counts in old growth.
Ask. The assistant does the geoprocessing. Every step is kept.
- Ask.Say what you want in plain language, to the assistant you already use.
- It runs real geoprocessing.The assistant calls NativeGIS tools over MCP: open a file, check its coordinate system, query features, process LiDAR, export a result. Results are exact geometry, not screenshots of a map.
- Every step is recorded.A log of each tool call, its parameters and its result. Read what was done, and replay it on new data.
- Outputs open in any GIS.Standard files. Exports never overwrite. Every response states its coordinate system and units.
Long jobs run in the background. The assistant starts them, checks progress and picks up the result when they finish.
The whole HJ Andrews Experimental Forest, on one desktop.
We took the public 2021 USGS 3DEP LiDAR for the HJ Andrews Experimental Forest in Oregon and ran the full forest chain over all of it.
| Area | 63.7 km² |
| LiDAR points | 3.57 billion |
| Machine | One desktop PC (8 cores, 64 GB), shared with other work the whole time |
| Time | 3 h 22 min, including a 45 min download of 25 GB of point data |
| Tiles | 1,114 processed, 0 failed |
| Trees found | 832,412 (131 per hectare) |
| Canopy cover | 94.0 percent |
| Mean canopy height | 25.1 m |
| Trees 50 m or taller | 103,077 |
- One desktop, one afternoon. A forest this size usually means a server, a cloud bill or a week of waiting.
- Tiling does not change the answer. Inside the test window, the tiled run found 8,409 trees against 8,410 for the single pass. Height RMSE between them is 0.02 m. A tree on a tile seam is counted once.
- Trees found are dominant and codominant tops, 5 m and up. Small trees under the canopy are under-counted.
The same pipeline ran on a 56 ha window first, where we checked it against independent software (see accuracy above).
Open the demo: the forest's maps, 2008 to 2021 canopy change, a stand in 3D you can turn, burn severity from the Lookout Fire, generated sampling plots, and the recorded session that replays all 87 steps.
People who need the layer more than the software.
Questions people ask first.
Does my data leave my machine?
No. NativeGIS runs on your computer and reads your files where they are. Usage counts are off by default, and nothing is sent anywhere.
What formats does it read and write?
It reads LAS/LAZ point clouds, GeoTIFF and COG rasters, Shapefile, GeoPackage, GeoJSON, KML/KMZ and CSV, and EPT point-cloud archives on disk. It writes vectors to Shapefile, GeoPackage, GeoJSON, KML/KMZ and CSV, rasters to Cloud Optimized GeoTIFF, and trees to GeoPackage, with each tree's points to LAS. Exports never overwrite.
How accurate is it?
We measure it against independent software and against other surveys, and we publish the results, including where it is weak. On a 56 ha forest window the terrain model is within 0.23 m RMSE of a denser survey, and the tree layer agrees with an independent run to 4.1 percent omission and commission. It does not stand in for field plots when you need small and suppressed trees.
Do I need to know GIS?
It helps to know what you want. You do not need to know which tool to open or in what order.
What does it cost?
Start with a free 3-month trial: confirm your email and every tool is yours to use. After that, NativeGIS is part of the NativeCAD MCP / Modeler seat, $200 per user per year, and the same key also opens NativeCAD and NativeBIM.
Which assistants work with it?
Any assistant that supports MCP over stdio.
Try NativeGIS on your own data.
NativeGIS is in early access. Tell us what data you work with and the layer you need, and we'll set you up.