AI Agent

An AI agent inside QGIS 4: it inspects your project and processes data from a plain-language request. It works as a loop — looks at the project, calls tools, sees the results, decides the next step. Changes are applied only after you confirm them.

What it does

Twelve domains, 65 tools:

Domain	Example requests
inspect	“what layers do I have?”, “what did I select?”, “show me the map”
project	“load this file”, “add an OSM basemap”, “load a table from PostGIS”
style	“make the rivers blue”, “colour by population”, “label with names”
processing	“build a 500 m buffer”, “clip by the district boundary”, “compute NDVI”
osm	“download the cafes in Tver”, “roads except unpaved ones from OSM”
edit	“fix the misspelled name”, “delete the features I selected”
fields	“add a virtual field with the area in hectares”, “rename nm to name”
layout	“make an A4 map sheet with a legend and export it to PDF”
python	the escape hatch: a PyQGIS snippet you read and approve first
web	“find the EPSG code”, “read this documentation page”, “geocode this place”
annotations	“mark this place”, “add a note to the map”, “remove that annotation”
three_d	“open a 3D view of this project”

On a vision-capable model the agent sees the map: it renders the canvas or a print layout to an image and judges colours, labels and composition by eye. And after you press Apply it verifies itself — re-reads the project, compares the result with what you asked for and queues fixes when something is off.

The safety model

You never watch the agent mutate your project unsupervised:

• Local reading tools (listing layers, describing fields, querying data) run immediately — they cannot change anything. Their results may still be sent to the configured model as part of the agent loop.
• Network reading tools wait in a plan for per-call confirmation. A web-only batch does not change or snapshot the QGIS project.
• Writing tools (styling, processing, loading data) are collected into a plan card. Nothing runs until you press Apply; Cancel discards the whole batch.

Arguments are validated at queue time, while the agent is still listening — so a metre buffer on a degree-based layer is rejected with a ready reprojection plan instead of producing garbage.

What it needs

• QGIS 4.0 or newer.
• An account and an API key with a language-model provider — any OpenAI-compatible endpoint or Anthropic. Local servers (Ollama, LM Studio) need no key at all.
• Sending a request shares your prompt, recent chat, project notes and basic project metadata with the configured endpoint. Sharing sensitive GIS data and tool results—feature attribute values, exact map and layer extents, layer filters and sources, style categories, Processing and Python results, and rendered map or layout images—is a separate option that is off by default. Local servers can still store or forward data. Read Data and privacy before opening a sensitive project.
• Each optional web call is confirmed separately. Search terms go to DuckDuckGo or, on fallback, Wikipedia; geocoding goes to the Photon demo or custom Nominatim service selected in Settings; and page reads contact the approved public HTTPS host. The public OSMF Nominatim instance is intentionally not offered. Geocoding starts disabled; Photon permits reasonable demo use but may throttle and has no uptime guarantee. Private hosts and credential-bearing URLs are rejected. Returned web content is untrusted data, can be forwarded to the chosen model and is not cached on disk. This flow is controlled by the per-call web confirmation, not by the sensitive-GIS-data option.
• API keys are stored encrypted in the QGIS authentication database, never in the project or in the QGIS settings file. The plugin has no external Python dependencies at all: install the archive and it runs.

Ready to try? Start with Setup, then see Usage for what to ask.

The interface

Available in English and Russian, following the QGIS interface language. The agent itself replies in whatever language you write to it.