It sees what you see
Before it answers “what am I looking at?”, it checks which named parts of the place are on screen and how much of the view each one fills. Tap anything with a name and the question writes itself.
The field.tours engine
An engine for AI agents inside real 3D places
The field.tours engine runs an AI agent inside a streamed 3D scan of a real place. It sees what the visitor sees, moves the camera, points at and lights up what it talks about, and answers only from the place's own cited sources. Every field.tours tour runs on it.
What it does
What can an AI agent do inside a 3D scan?
Six things, all working in the tours today.
“What am I looking at?”
“Show me where the reading begins”
It takes you there
Ask to see something and the camera flies there: to a stop on the route, or to any named part of the place, even where no stop goes. The visitor leads; the route is a suggestion.
It points at the real thing
It lights up the building, the rune or the crevasse field it is talking about, on the 3D model itself. Not a pin, not a pop-up box.
It answers from sources
Every answer comes from texts and sources written by the people who know the place. The source is one tap away, and when something is not in them, the agent says so.
It can have a body
In a Custom story the agent is a character in the scene, such as a survey drone over a glacier, with actions of its own that the visitor sees it perform.
It reads live records
A tour can be connected to a museum's own catalogue. The agent reads the live record, so a correction made at the museum reaches the tour on the next visit.
Sources
Where do its answers come from?
Every part of the place that matters gets a name, a sentence about it and its sources, written by a curator, a researcher or an engineer. The agent works from those names and texts, and from nothing else.
We do not let the agent guess what something is from the pixels. A guess cannot say where it came from, and for a museum, a school or a power company an answer without a source is not an answer.
When sources disagree, the tour shows both instead of quietly picking one. Asked about something outside its material, the agent says that it is not in its sources.
That is also what separates a field.tours place from a generated 3D world: the place is measured, scanned from the real thing, and every named part of it is backed by someone who knows it.
Where it runs
What does it run on?
The places are real scans, and the same tour runs on every screen a visitor meets.
Real scans, streamed
3D Tiles, point clouds, photogrammetry, BIM/IFC and CAD models, in the billion-point class. The scan streams in tiles as the visitor moves, so a model far too big to download opens in seconds, with nothing to install.
One build, every screen
A browser tab on a phone or a computer, full screen on a touch display in kiosk mode, and a VR headset. Standard tours draw with WebGPU where the browser has it and WebGL everywhere else.
Two languages, typed or spoken
Swedish and English from the start, more on request. Visitors type or speak, and the answers suit the audience you choose: school classes, the public or specialists.
For everyone
Keyboard, screen readers and reduced motion work in every tour, because they are built into the engine rather than added per tour.
Two ways to use it
A Standard tour or a Custom story
Standard tourBuilt from your model and your texts. You manage it.
A streamed 3D model, numbered stops and your texts, with the agent answering from your sources. Stops, texts, name tags, sources and the agent's persona are content, not code, changed without a developer.
Custom storyWe write and build it with you.
Everything a Standard tour has, plus characters with a body and a voice, animation, simulation and bespoke interaction. We write, build and deploy it; the texts stay yours to change afterwards.
Studios, integrators and developers who want to build on the engine: tell us what you have in mind
Compared
How is it different from other ways to put AI in 3D?
Other approachesThe field.tours engine
AI characters built for game enginesAn agent inside a scan of a real place, in a browser
A chatbot next to a 3D viewerAn agent that moves the camera and lights up the model
An audio guide on the visitor's phoneA guide that knows what is on the screen in front of you
A world generated by AIA measured place, every named part backed by people who know it
Answers from whatever the model learnedAnswers only from cited sources, one tap away
Who makes it
Made in Stockholm by IVAR Studios
field.tours is made by IVAR Studios, the Stockholm studio of Martin Edström, director, and Fredrik Edström, producer, founded in 2016. More about the people behind it is on IVAR's site.
Questions
The questions we hear first
Is the field.tours host an AI agent?
Yes. It is a language model that acts in the 3D scene as well as talking: it moves the camera, points at and lights up parts of the model, checks what is on screen and looks up the place's own material before it answers. It acts only on the place it is in, and only on what the visitor asks.
Does it recognise objects in the scan by itself?
No, on purpose. People who know the place name its parts and write down what they are, with sources. The agent works with those names, so it can always say where an answer comes from. It never guesses what something is from the pixels.
Which language model does it use? Can we use our own?
Today Mistral Small through OpenRouter, with OpenAI's gpt-4o-mini as a fallback, as our privacy notice says. The engine is not tied to one model and has changed models without changes to the tours. Ask us about yours.
Can it work on our BIM model or digital twin?
Yes. It works on scans, point clouds, photogrammetry, BIM/IFC and CAD models, which we convert and stream. An industrial site can be shown on your own material, with your technical documentation as the sources.
Does it run on touch screens and in VR?
Yes. The same tour runs in a browser, full screen on a touch display in kiosk mode, and in a VR headset, with nothing to install.
How do you stop it making things up?
It answers only from the texts and sources written for the place, cites them, and says that something is not in its sources rather than guess. Conflicting sources are shown, not quietly resolved.
Can we build on the engine ourselves?
Talk to us. We are looking for pilot partners and scope each first project together, and we want to hear from studios and developers who have a place, a story or a product in mind.
What does it cost?
There is no price list. We plan a first project together with each organisation and shape the offer with our pilot partners.
See it live
Watch the agent work in a real place.
The demos show everything on this page: a runestone, a farm at Skansen, a ship from 1759 and a glacier front.