Example: surplus solar power for the car
One useful application is charging the car with available solar energy. The inverter supplies the current surplus power directly as a single value. So Hermes does not have to do the math with the home's production and consumption.
Our EV charger is not in Home Assistant either. Hermes reads the inverter's value. It controls the charger through the charger's own local API. Home Assistant does not sit between those two connections.
The rule behind this is deliberately simple. Charging starts only after enough surplus has been available for long enough. If the surplus drops or a device stops responding, the automation stops safely. A change to the charging rule needs confirmation first. Meanwhile, you can ask through chat how much surplus is available and whether the car is charging. You can also ask why the automation did or did not act.
Chat makes separate integrations usable
Without chat, these are separate dashboards and technical connections. With Hermes, you can request the same information in ordinary language. You do not need to remember where to find a status or which application belongs to which device.
That does not mean a language model decides freely. Hermes translates the question into a known check or action. Repeatable tasks remain fixed rules with clear boundaries. The agent can explain, but the automation decides when to control a device.
That separation makes the system more predictable. You can phrase a question freely. The execution stays fixed.
Local-first without forcing everything to stay local
Sisu's base runs at home on the Pi: Home Assistant and the device connections stay close to the source. We do use an external service for the language model and chat channel. Sisu is therefore a hybrid, local-first setup.
That is a practical choice. Local control does not need to pass through a manufacturer's cloud. If the internet goes down, local rules can keep working, but chat may be temporarily unavailable.
The term local-first comes from the essay Local-first software by Ink & Switch, published in 2019. In local-first software, the data on your own device is the main copy. Servers only hold an extra copy. We apply that idea to controlling the house. Read more about what stays local in a hybrid AI setup. To see whether a setup like this fits your business, read local AI vs. cloud AI.
Keep access simple and limited
Local does not automatically mean secure. We keep administration interfaces private and use separate accounts with limited permissions. We give each integration access only to what it needs.
For actions with a physical effect, we start with read-only access. Then we test one targeted action. Only when the behavior is clear do we add automation. If data is missing or unreliable, nothing happens. Anyone who wants to change an important rule has to confirm it deliberately first. Our AI safety checklist for business data applies the same rules to a business.
How to build the same foundation
Start with a Raspberry Pi 5, reliable storage such as an NVMe SSD, and Raspberry Pi OS.
Install Hermes and Home Assistant Container on the same machine, but keep their roles separate.
Connect Home Assistant through the official Hermes plugin with a limited account. Add only the devices that fit it well.
Use a direct local API when it is simpler and clearer for a device.
Add one chat channel and start with questions that only retrieve information.
Then automate one narrow task and also test what happens when data is missing.
A first version does not need more than that. A small working system teaches you where the value is faster than a house full of unfinished integrations.
Is a Raspberry Pi 5 powerful enough?
For this role, yes. Our Raspberry Pi 5 has 8 GB of RAM and a 1 TB NVMe SSD. That is enough for the connections, Home Assistant, and lightweight automation.
The large language model does not need to run on the device itself. That keeps the local machine compact and energy efficient. A Pi does not replace heavy AI hardware or a developer workstation. It does not need to. Sisu is the control layer that is always ready at home. Heavy computing can happen elsewhere. For heavier local models, see what computer you need to run AI locally.
What does this home setup teach a business?
Let a local agent work close to the processes. Choose the simplest reliable connection for each system, and grant only the permissions needed. A business does not need a Raspberry Pi for this; what matters is that division of work.
A business environment needs more work on user administration, logging, updates, and support. AI backup and recovery is part of that too: encrypted, stored at a separate location in the customer's network, never on the node. The principle stays the same. You do not have to move everything to one cloud platform to make AI useful.
On the page about local AI for businesses, you can read how we translate this into a managed installation. Hermes as an AI agent shows which tasks the agent can carry out. Under security and privacy, we explain how we limit access and data flows. Voltti supports pilots in Belgium and the Netherlands.
For the business hardware decision, read when you need a Linux computer for an AI agent.
