My son is in his Lego era. I mean, I don’t think most folks ever leave their Lego era, but he’s deep in the thick of his. His brain sees a box of assorted Lego bricks and sees the potential: there’s a rocketship, there’s a race car, there’s a soccer stadium. The raw ingredients present in a box of random pieces mean that anything is possible. I get it.
This week, the tech world is all abuzz thanks to OpenClaw (formerly known as Moltbot (formerly known as Clawdbot)). And for good reason. Peter’s invention takes the incredible AI coding agent superpower many of us have become accustomed to (Claude Code, Codex, whichever your preference), layers on some connectivity tools (so you can message it from anywhere), and emphasizes on-device memory. I was immediately in the “I’ve got to try this” subset of the tech sphere.
I didn’t rush out to buy a Mac mini, because I’ve had one operating in a similar vein for a while. The idea of an always-on, fully-capable machine (that also happens to run macOS instead of a Linux box somewhere in the cloud) has been my comfort zone for quite a while because of the capabilities it enables. So adding OpenClaw to the mix seemed like the obvious solution. The trick - as so many it seems are asking - is deciding what to use it for. Like a box of unsorted Legos, what should you build with it?
I had some ideas. Three, at least.
Before we get there, I want to give a mini review of OpenClaw from days of solid experimentation:
- This is new, constantly changing stuff. I wouldn’t recommend this (at the moment) for most people. If you’ve never run Claude Code with –dangerously-skip-permissions or if you don’t know what that means… I would probably sit this out for a bit, until it becomes less of a Wild West.
- There are security risks. Google “clawdbot” (or either of the two newer names) and I’m sure you’ll have lots of evidence to prove this. Just because you can do something doesn’t mean you should or it wouldn’t result in someone somewhere seeing the contents of your desktop, sending emails on your behalf, or any sort of nefarious shenanigans script kiddies can get up to.
- The biggest weakness for me is that OpenClaw uses a ton of tokens. You can exhaust your Claude or ChatGPT subscriptions pretty quickly (ask me how I know) or if you give it an API key for Anthropic or OpenAI, you might want to budget this alongside “vacations” or “mortgage”. (Slight exaggeration but it’s a bit nuts.)
- All that said, the ability I now have to chat (on my phone, via Telegram) with a bot who can read or modify files on my always-on Mac mini, can run programs and send me the output, can access my calendar, or do - practically speaking - anything a person using a computer can do, provided it has the right permissions is fairly incredible.
- It should go without saying that none of this is required. You can still be a productive human without a virtual lobster-inspired robot assisting you. At least for now.
Where was I? Oh yes. If you have the tools to make anything happen… what would you make happen? Here’s what I did:
Use case #1: a persistent daily journal
You’ve got your Day Ones and your Bears and your Obsidian daily notes and a hundred other tools to keep track of what you did and what you need to do. Why not throw another on the pile? Unfortunately for all the previously mentioned apps, I’m not great at any of them. Between the day job, the side projects, a crazy busy family, and everything else… time is at a premium for me. In fact, I’m halfway through my second attempt at an app to solve this for me. After getting OpenClaw set up, I abandoned that project.
If I’m in a meeting and I need to write down something, I can message it to my bot. If I think of something I need to do tomorrow, but it’s 11:00pm and I’m half-asleep, I can message it to my bot. If I think of a random thing, “hey, when I write this OpenClaw blog post, I should mention how I have three use cases at the moment”, I can tell it to my bot.
Then! Whenever I want, I can say “hey, remind me of what I should write in my blog post?” and it knows what I told it. Nothing groundbreaking. It’s how brains work. But it saves mine from having to remember everything.
I also have it sum up the things I’ve said to remember or things I need to do and message it to me in the morning. Nothing fancy, but now it’s not hard to do. As a bonus, I also have it create Apple Reminders for me when I want (for only the things I want). Again, not groundbreaking, but as an aspirational Reminders user, it helps me bridge that gap.
Use case #2: an email sorter
Remember when I said I already do some tasks with my always-on device? One of which was inspired by Zach Gage who shared a tool a year or two ago to filter and sort his email. I set it up not long after he shared it and it’s been running locally using a local LLM model for a long while now.
The rationale here is that email is the worst, but it’s also annoyingly important. With all the aforementioned busyness, I have emails from school, kid activities, financial notifications, actual written-by-human emails occasionally, and trying to keep an inbox clean is a fool’s errand. So with Email Hole, I had a helpful but primitive robot running locally and sorting my email into the right folders. This has been good… but what if it could be better?
I mentioned earlier that OpenClaw combines a lot of other Peter Steinberger tools and one that I hadn’t known about, gog, is perfect for interacting with my Google-based email accounts. But also - in my quest to not pay an arm and a leg for LLM tokens for OpenClaw - I went down the path of using the open source gpt-oss model via LM Studio. It technically works, and it’s free, but it’s far from the speed or smarts of Opus, Sonnet, etc. But, for things like categorizing emails behind the scenes, it’s perfect.
So I’ve updated my script to use gog to connect to my Gmail and use gpt-oss instead of the previous ollama model I’d been using. It’s already running every 5 minutes. None of this is OpenClaw-specific.
However… the sorter has a config file that tells the local LLM what types of emails should be in what folders. Now any time I have a categorization I want to change—maybe the school started using a new newsletter provider and I want anything from that domain to go to the school folder—I just Telegram (are we using that as a verb?) OpenClaw and it updates the AI rules for the email sorter. The magic of being able to write to files!
Use case #3: a robot product manager
I’m always looking for ways to improve my main side project, Sticky Widgets, especially in the non-code department. (Claude Code has that covered.) I recently released a command line tool, xcstrings-localizer, to make LLM-based translations easy. But I have a lot of ideas for the pre-development portion of running a one-person app.
The first thing I started with was analyzing user feedback. I have an in-app feedback form, using tally.so, which fortunately has an API. I also wanted to consider App Store reviews, which fortunately has a public feed as well.
One of the OpenClaw-initiated realizations for me was that command line interfaces are so useful for me to trigger manually, for a “standard” agent like Claude Code, and now from OpenClaw. CLIs are becoming my first step when I want to build something.
So, I built a tool to fetch these sources of feedback, to pull from a basic ABOUT.md which provides a summary of Sticky Widgets and its features, and ask gpt-oss running locally to learn about my app, take the feedback, and analyze what users are saying (emphasizing frequency, recency, and level of impact). (ABOUT.md was generated when I asked Claude Code to describe the app after analyzing my codebase as well as publicly available information about the app. I plan on automating this as I make changes to the app as well.)
I also added a feature to allow me to pass any arbitrary question as a supplemental prompt for the LLM to answer about the feedback data it had access to. The CLI is fine and operatable when I trigger it manually, but it shines when I asked OpenClaw to ask it something. I also schedule this to give me weekly updates.
This is just the start and I’m currently working on an additional layer that will create GitHub issues (via the gh CLI) or supplement existing ones with fresh feedback from users. GitHub (and its comprehensive command line tool) is becoming my source of truth for Sticky Widgets product planning. I have many more ideas on product-side automation of Sticky Widgets, especially around piping in download, subscription, and usage metrics. I’m only scratching the surface here.
OpenClaw is fascinating as a product and tool, but its primary use for me is being the connective tissue that makes interacting with tools of my own creation easier. Its biggest innovation is making me realize that so many things are possible. It’s an odd feeling, but the “I wonder if I can…” questions I ask myself often have a resounding “yes” answer these days. It’s never been easier to build what I want.