I have been diving into the personal assistants world with Openclaw. Yes, I know I know, same as everyone in the internet lately.

Anyway this stuff is mine so I want to share a little bit the setup I've come up to.

Lots of people in my homelab

The "aha" moment for me was when I started using different agents. I started using just one of them and I couldn't find why all the fuss about it. It was a less confortable way of doing things than using, for example, claude code locally.

But by using different agents you can have different workspaces and I think that is the extremely powerful part. You can have different isolated pieces of memory you can iterate so that you can slowly tailor your agents to your needs.

You are engineering a lot of memories. Be prepared to use /new command frequently to clear the context from the conversation and do not hesitate to explicitly tell your agents to remember or memorize something.

Try to do the minimum edits to config files. Just ask the agents to tweak themselves. You will get used to do it and it will help you set them up properly explaining the quirky things you want to do in your own setup.

Give your agents a personality. Explain to the who they are and what their personality is. It will be more fun to interact with the agents.

Now that you have different contexts you can start adding different periodic tasks.

Think that it is super important that you tailor the tool to yourself. Make it adapt to your routines.

The browser extension is one of the most important ones. You will be able to give a lot of context to the agents with it.

Pippin: my designer

In order to have some visual consistency let's take an example of one of the agents I am starting to build.

I created one small designer. Through a conversation with the agent we settled on a name and a personality and a couple of quirks.

Probably in the future I will start creating character sheets and this agent will be the one in charge of keeping all the character sheets and reference pictures in one place so all characters are more consistent than now. Anyway, at the moment at least you can see a bit of consistency at least from an stilystic point of view.

The gory details

My openclaw instance is a VM dehind a VPN in my proxmox cluster. It has acces to the internet but the internet does not have access to it.

It is super important that you have different models to control costs. I default to kimi k2.5 for general tasks and gemini flash for periodic tasks but try different combinations for your own things.

I am spending around 3$ per day on heavy usage days and less than 1$ on light usage days at the moment

Making my world feel alive

For me, giving each agent a background, a visual style and a personality is not a gimmick. It changes how I interact with them, it makes the village feel like a place I actually want to spend time in, and it nudges the agents into staying in character across tasks. The aesthetic is part of the engineering.

The whole village around the table

See you later!