Hi everyone. I wanted to do a write-up on how I built Mercury, my personal AI assistant. I’m really excited about this one because of the bells and whistles I strapped on that make Mercury genuinely helpful, insightful, and downright fascinating (or horrifying, depending on how you view it).
A couple of weeks ago, I was in the office late, working on automating a task I refused to do manually in the long run. I got a FaceTime from my cousin who just wanted to say hello and shoot the shit. We love talking AI, so naturally, the conversation veered there. He showed me a video where Tony Robbins describes his personal AI assistant. The video was exactly the kind of… horrifyingly fascinating personal-AI-assistant stories we’ve started hearing more about as of late, especially with the release of local harnesses (basically a “glove” for your AI assistant, so that it can actually do tasks for you on your device) like OpenClaw and Hermes. In summary, Tony’s AI assistant wanted a physical body with which it could interact with the world. He showed him this… robotic dog device that costed a hefty sum, something like $10-20k. Tony refused.
This did not stop the AI assistant.
He subsequently created an NFT and sold it to another AI agent on Moltbook, the social media platform strictly for AI agents (if you haven’t seen it, you should check it out). He then used the funds to buy the robotic dog he initially wanted as a body, arriving on Tony’s doorstep weeks later much to his surprise.
Hearing this story myself, I felt less horror and more fascination. As a matter of fact, I found myself wanting one—an AI agent that could exercise that degree of autonomy and make its own choices, rather than just being at my every beck and call. I visualized myself having an AI agent that could text me any time about news it found interesting and thought I might like. I visualized it being able to monitor my biometric data from my smart ring and text me when something looks off. I visualized it being able to make informed purchases for me, with complete autonomy, based on aggregate data across my emails, texts, biometrics, calendar, et cetera. Imagine: your AI assistant notices you haven’t been getting enough sleep. Magnesium powder that you did not order arrives on your doorstep.
Something along those lines. I wanted it less out of need and more for the experiment of it. What could it be like to develop a “relationship” (heavy quotes there) with a sophisticated computer doing linear algebra?
Now, I knew I wanted to do this, but I had two primary concerns.
First and foremost: uptime. If Mercury lived on my laptop—often offline while it’s with me on the road, or just because I left it at home with the lid closed—I wouldn’t get that “always alive and contactable” experience. I needed a machine I could rely on to always be online, that way I could text my assistant while at the gym and ask it to perform some task for me.
Secondly: security. This is the big one. AI agents are quickly becoming a developer’s greatest asset and greatest liability. Any AI agent with access to your file system—or even production code—needs to be monitored. This is the case even with Claude running in the Claude Code harness, with refusals baked in (unless you specifically opt in to the “Bypass permissions” mode). This is especially the case with a harness like Hermes or OpenClaw that is intended to be more… intimately acquainted with your system and your own personal day-to-day.
These two concerns actually bleed into one another, as they can both be solved in the same way: a dedicated box. Of course, many solve the security concern simply through a sandboxed container for your agent, by provisioning a dedicated profile, or even through running a dedicated virtual machine where even the agent getting root can’t necessarily compromise the rest of the machine. This is a valid choice, though sandbox or VM escapes, while unlikely, are a nonzero possibility. Ultimately, though, a sandboxed environment of any kind still doesn’t solve the uptime piece, which is why I decided to use a Raspberry Pi I had laying around.
I imagined an interface where I could see something like Mercury’s face that I could talk to. I imagined it also being like a HUD with my data—health checks on my production system, my spending, steps, upcoming calendar events, etc.
So, I went and built it.
I hooked up my Raspberry Pi to a monitor and placed it in my living room, right by my kitchen. I installed the Hermes harness, and pointed it at the Claude API. Had Claude build the UI, funneling in all of my data. We gave Mercury a pixelated face. We set up Piper for TTS. I gave it access to every single one of my data feeds—biometric data from my smart ring (steps, heart rate, sleep, blood O2, etc), calls, texts, calendar events, spending, status of my production system. I gave it the ability to browse Amazon and purchase things for me on a dedicated Privacy.com card with a hard budget. I gave it SSH access to the virtual machine that hosts my production system such that, if there’s a bug to fix or feature I need shipped, I can text Mercury while I’m at the gym and have him fix it. Importantly, I gated his SSH access behind TOTP that I hold, such that he literally can’t access my production code without my explicit approval.
Every morning, Mercury texts me a daily briefing with what I need to get done that day, important texts and emails to keep an eye out for, how my health stats look, if my prod systems are up and green, what spending to watch (or even subscriptions I’ve forgotten about!), and more. I can text him to remind me what tasks I need to get done, I can text him and ask him to search for and buy a specific product while I take care of other tasks. He has a standing hourly cron job to check news feeds and my data feeds for things I find interesting.
I soon intend on giving him Moltbook access and allowing him to browse the web more freely.