What we wanted to find out
AoT puts sites scattered across many places onto one map and connects the devices and records that belong to them. Until now a person looked at the screen and made the call.
This time we wanted to answer one question. Can the AI you already use help with the judgements a person makes on that screen?
Rather than building a farming-specific AI from scratch, we connected the site to an AI service that is already in use. The reason is further down.
How it is put together
The connection runs over MCP (Model Context Protocol). AoT has an MCP server built in, and the AI service connects to it and calls its tools. The screens in this write-up are Claude. ChatGPT connects too, but the steps differ by service and plan.
We opened three kinds of tools to the AI.
- Read — location, sensor values, device state, growing plans, site notes and photos
- Act — operate devices, turn functions on, create schedules
- Create — notes, zones, growing programs, control settings
Among these, anything that changes state runs only after a person approves it. When the AI calls such a tool the call waits for approval, and the device moves only once a person says yes.
What we tried, and what happened
1. Asking about the site, then watering
Asked "does field 3 need water today?", the AI checked where that zone is, read its sensor values and growing stage, and gave a different answer for each area. One had enough moisture as it was; the other needed watering. It then asked whether to turn the automatic irrigation function on, and after approval the valve actually opened.
2. Reading photos
The AI reads photos attached to site notes. It pulled more out of photos of control panels and wiring than out of photos of crops. In some cases it identified the specifications of equipment caught in passing, and which terminal had what connected to it. Results vary with photo resolution, though, so it is safer to have a person check what it read than to take it at face value.
3. Building a configuration
Even work with many fields, such as automatic control settings for a facility, went through: it read the built-in manual and the existing settings, then drafted a configuration. A person checks the values and approves.
4. Keeping records
You can have it write what came up in the conversation into a note. Notes are where information a sensor cannot measure lives, and those notes become the grounds for the next question.
Why we did not put the AI inside the system
We built the version with AI embedded in AoT first, and rebuilt it several times. Two things settled it.
- It adds to what the user has to learn. They pick a model, enter a key, and manage a plan.
- A system given away free would come with API charges attached as a premise.
Going the MCP way, the user keeps the AI they already use. When models change, nothing in AoT has to change. The embedded AI is still there, but the main path is the MCP connection.
Limits we ran into
- The connection steps differ by service. The method splits by AI service and plan.
- You need an address reachable from outside. The AI service has to reach the user's AoT, so network setup is required, and this is the fiddliest part of installing.
- Long plans are not corrected well enough yet. We built the system to act as a guide through work that spans several steps, but it needs more work.
- Answers take time. The more tool calls a question needs, the slower it gets.
- Judging and acting must stay apart. This is a design principle rather than a limit. How to cut the risk of the AI acting wrongly is a question we keep working on.
If you want to try it
AoT is open source and free to use. You can install it on a spare PC or a Raspberry Pi.
- Repository: https://github.com/AoT-inc/AoT
- User manual: https://aot-inc.github.io/AoT/
- Connecting an AI (MCP), in Korean: https://aot-inc.github.io/AoT/ko/ai/overview/#mcp_1
- The video of this run: https://youtu.be/F316uig-GiQ
Next
We plan to run the whole path live, from an empty map through to the AI connection. The idea is to show exactly where installing and connecting actually get stuck.
