I'd poked at this before building it properly: n8n workflows connected to my Google Calendar, Gemini's built-in Google Suite integration as an assistant. Both worked in a limited way. n8n boxed me into its workflow logic and needed a paid VPS to stay running. Gemini's integration was tighter but limited to whatever Google handed it.
The bigger reason was privacy. Having Claude, ChatGPT, or Gemini sit on top of my calendar, my data, and the things I ask is a real exposure, and a Pro subscription doesn't change that. If I was going to give an assistant access to my calendar, email, and trading research, I wanted it running on hardware I controlled. If it could also reference my own trading rules, there was a chance it could pay for itself.
The harness
OpenClaw is an open-source AI harness, a prebuilt agent shell that needs a brain plugged in through a local LLM. It ships with a "soul" file for the agent's personality and purpose, a skills section for what it can reference and act on, and a memory area for markdown files and context. Local models matter more in this community than most because it started with people attaching OpenClaw to API access that was effectively unlimited at a flat price, until that access got restricted. Nobody can throttle a model sitting on your own hardware.
Buying the hardware
My work machine was off the table; I wasn't going to risk it for a side project. A dedicated PC build was expensive, mostly RAM prices. An old PC couldn't keep up once I checked the specs. I landed on a Mac Mini: 16GB unified memory, M4 chip, Unix-based, around $600.
Picking a model
Set up and connected to Discord, I tested models through Ollama: Qwen, Gemma, Qwen Coder. Most of it was learning how hardware and parameter size trade off, how fast a model responds, and how good the answers are past demo-level questions. The models differed almost like separate personalities in how they phrased things and what they defaulted to when a question was ambiguous.
Skills and memory
Once it was reachable from my phone, the work was memory and context. ClawHub, an open-source repository of skills as markdown files built by others running the same setup, taught me as much about prompt and skill design as anything else in the project. Memory took longer. The early version dumped information into markdown files with no thought to how the agent would find anything. I rearchitected it more like a graph than a flat folder: files referencing other files, so the agent has a path to the right context instead of scanning everything.
What it costs
"Free once you've bought the hardware" is half true. Anything needing live data (search, web access, Google) runs through third-party APIs that cost money. I use OpenRouter to switch providers and contain that. Open models are also still behind Claude and ChatGPT on a lot of tasks, and I'm fine using a hosted API for what local models aren't good at yet. This was never about ideological purity, just about not wanting my data on someone else's server by default.
The thesis
Under the personal-assistant project is a bet: we're moving toward a "you'll own nothing" economy of subscriptions instead of ownership and "free" apps that sell you to data brokers. I think privacy will matter to more people over time, and a self-hosted AI assistant, like running your own server for photos instead of trusting a cloud service, is a niche move that's growing. I built this to test that on myself first. If it holds up, the skills here are also the skills behind offering something like it as a service.