A friend recently sent me a project for an ESP32 MCP server. The repository is not new, but as someone much more comfortable with software than hardware, I found the idea immediately interesting.
The first obvious use case is asking for data from an ESP32 in plain prompts: “What is the temperature in the greenhouse right now?” or “How moist is the soil?”
But a question quickly followed: why bring an LLM and MCP into this at all when reading sensor values can be done far more simply?
For some reason, I could not let that thought go. During yet another not-especially-important meeting, I wondered: what if I went in the other direction and gave a local model the job of control, rather than data retrieval?
I recently became the owner of a country house, and, to put it mildly, I do not know much about growing plants. The idea of handing a local model the task of analysing a greenhouse’s condition quickly took hold.
How it could work
The initial concept is fairly simple.
The model periodically receives temperature, air humidity, soil moisture, and other measurements. It has a goal — say, to create the best possible conditions for growing tomatoes.
It then analyses the situation, chooses an appropriate action, and uses the ESP32 to control irrigation, ventilation, and other systems.
Instead of a huge collection of prewritten rules, ideally we tell the system what outcome we want and let it try to maintain the right conditions. Naturally, with limits, safeguards, and the ability to intervene manually.
But the idea quickly grew beyond sensors alone.
More than a smart greenhouse
The next stage brought a longer list:
- video monitoring and plant-condition analysis;
- detecting signs of disease;
- analysing soil and recommending fertiliser;
- pest detection;
- connecting a knowledge base on botany and horticulture;
- RAG and perhaps eventually further training the model on agronomy data.
This is no longer just a “smart greenhouse.” It is a local AI agronomist that constantly observes the plants and tries to make its own decisions within a defined goal.
To test the hypothesis rather than merely the technical implementation, I told a friend about it. He studied at the Kostroma State Agricultural Academy and clearly knows far more than I do about plants, greenhouses, and tomatoes.
So I was a little surprised when, instead of “you have come up with some nonsense again,” I heard something along these lines: the idea is entirely feasible; it just needs testing in a real greenhouse.
It looks as though the next year will be all about the AI garden. If you are interested, I will share the process, discoveries, mistakes, and, of course, the final results as the project moves forward.
Thank you for reading.
Your Quackmaker 🦆