GeoAgent is a QGIS plugin that lets you run geospatial workflows in plain English. Ask it to load data, select features, or chain geoprocessing steps — it finds the right QGIS algorithms, fills in the parameters, and adds the results to your map.
Install from QGIS Plugin Repository · Report an issue
- Layer management — add vector/raster layers from files or URLs, list layers and columns, zoom, remove
- Selection & queries — select features by attribute (
=,>,contains, ...) or by geometry (largest, inside, intersecting, ...) - Geoprocessing in natural language — buffer, clip, dissolve, zonal statistics, and any other algorithm from the QGIS processing registry (native, GDAL, GRASS, ...)
- Multi-step workflows — one request like "add demo.shp, buffer it by 5 km, then clip raster.tif with the buffer" is decomposed into ordered tasks, with outputs passed from one step to the next and automatic retry on failures
- Your choice of LLM — Ollama (local, free), OpenAI, Google Gemini, or Anthropic Claude
The plugin has two modes, selectable with the radio buttons in the panel:
| Mode | What it's for | Example |
|---|---|---|
| General (default) | Questions, data exploration, selections, layer management | "Which layers do I have?" |
| Processing | Running geoprocessing algorithms, including multi-step chains | "Buffer cities by 1 km and clip roads with it" |
From the QGIS Plugin Repository (recommended): Plugins → Manage and Install Plugins, search for GeoAgent, click Install Plugin.
Manual: download the latest release, extract it into your QGIS plugins folder, restart QGIS, and enable the plugin.
| OS | Plugins folder |
|---|---|
| Windows | %APPDATA%\QGIS\QGIS3\profiles\default\python\plugins\ |
| macOS | ~/Library/Application Support/QGIS/QGIS3/profiles/default/python/plugins/ |
| Linux | ~/.local/share/QGIS/QGIS3/profiles/default/python/plugins/ |
Requires QGIS 3.2 or newer (QGIS 4 supported). On first use, open the plugin's Settings tab and click Install dependencies — this installs the required Python packages (LangGraph, LangChain provider libraries) into your QGIS Python environment.
Pick one provider in the dropdown, then:
| Provider | API key | Example models |
|---|---|---|
| Ollama | not needed (runs locally) | llama3.2:3b, llama3.1:8b |
| OpenAI | platform.openai.com | gpt-5, gpt-5-mini |
| Gemini | aistudio.google.com | gemini-3-flash-preview, gemini-2.5-pro |
| Anthropic | platform.claude.com | claude-sonnet-5, claude-haiku-4-5 |
For the API-key providers, paste your key, type a model name, and click Save settings (settings survive QGIS restarts).
Ollama quick start (free, no account, runs on your machine):
# 1. Install from https://ollama.com/download, then:
ollama pull llama3.2:3b # download a model
ollama serve # make sure the server is runningKeep the default base URL (http://localhost:11434) and model name in the plugin, or change them to match your setup. If the model isn't installed yet, GeoAgent offers to pull it for you.
Tip: small local models work well for General mode, but Processing mode involves structured multi-step reasoning — a larger local model (8B+) or a cloud provider gives noticeably better results.
"Add a shapefile from C:\data\cities.shp"
"What columns does the cities layer have?"
"Select cities with population greater than 100000"
"Zoom to the boundary layer"
Single operations:
"Create a 500 m buffer around the cities layer"
"Clip the roads layer by the study area"
"Dissolve the parcels by district field"
Or a whole workflow in one message:
"Add C:\data\demo.shp, create a 5 km buffer for each shape,
and clip raster.tif with the buffered layer"
Behind the scenes, GeoAgent:
- Splits your request into ordered tasks and decides which need geoprocessing
- Searches the full QGIS algorithm registry and picks the best match per task
- Fills the algorithm parameters from your wording (units like "5 km" are converted)
- Runs the algorithm; on failure it diagnoses the error and retries with a fix (up to 2 times)
- Loads outputs into your project and feeds them into the next task
- Posts a summary of what was done
- Name your layers and include units: "buffer cities by 500 m", not "make a buffer"
- Reference earlier results naturally: "...and clip the roads with that buffer"
- File paths work as inputs even if the layer isn't loaded yet
Temperature (0–1, default 0.8) — lower is more deterministic, better for precise processing tasks. Max tokens — caps response length. Export chat / Clear chat — save or reset the conversation. The Logs tab shows each workflow step, which is the first place to look when something goes wrong.
| Problem | Fix |
|---|---|
| Ollama connection failed | Check ollama serve is running and the base URL is http://localhost:11434 |
| Model not responding | Verify the API key and model name; check your internet connection |
| "Layer not found" | Ask "what layers are loaded?" and use the exact name (or the file path) |
| Wrong algorithm/parameters chosen | Rephrase with the operation and units stated explicitly; check the Logs tab to see what was selected |
| Missing package errors | Settings tab → Install dependencies, then restart QGIS |
| Plugin misbehaving after update | Restart QGIS, or disable and re-enable the plugin |
geo_agent/
├── agents/ # LangGraph graphs
│ ├── graph.py # entry points (general + processing modes)
│ ├── workflow.py # multi-step task loop (decompose → route → execute)
│ ├── geoprocessing_flow.py# geoprocessing sub-graph (discover → params → run → retry)
│ ├── states.py # graph state definitions
│ └── schemas.py # structured-output schemas
├── tools/ # LangChain tools wrapping QGIS (I/O, selection, processing)
├── llm/ # Provider clients + background worker thread
├── prompts/ # System prompts
├── dialogs/ # Qt UI
├── config/ # Settings and constants
├── utils/ # Dependency installer, layer matching, helpers
└── geo_agent.py # Main plugin class
Contributions are welcome — fork, create a feature branch, and open a pull request. For bugs, open an issue with your QGIS version, plugin version, and steps to reproduce.
MIT License — see LICENSE. GeoAgent's own code is MIT-licensed; when it runs inside QGIS (GPL v2+, with PyQt under GPL v3), the combined work is governed by those licenses' terms, as with any QGIS plugin.
Authors: Tek Kshetri, Rabin Ojha
Related projects that inspired this work: QChatGPT, GeoAI, QGIS MCP. Thanks to the QGIS, LangChain, and open-source GIS communities.