The Hottest AI Tool of 2026 Is One Your Firm Shouldn't Touch Yet
150,000 developers are building AI agents right now. Your future competitor is probably one of them.
Image created by Nano Banana Pro
I Built My Own AI Agent Army With OpenClaw. Here’s What Happened.
TL;DR: OpenClaw is the first truly agentic AI tool that runs on your own hardware and actually does things on your behalf. I bought a Mac Mini, set it up, and built a team of specialized AI agents that research, draft content, manage my CRM, and brief me every morning. It’s real, it’s powerful, and it’s not ready for law firms. But it’s a preview of where everything is headed.
Two weeks ago, if you worked anywhere near technology, you couldn’t escape one name: OpenClaw.
It was on every podcast. Every LinkedIn feed. Every Slack channel I’m in. I’ve been doing this long enough to know that most “next big thing” announcements are noise. This one isn’t. If anything, the hype might be underselling what’s happening here.
Let me back up. Because the story behind this is almost as wild as the technology.
The Accidental Revolution of Peter Steinberger
Peter Steinberger is an Austrian software developer who spent 13 years building a company called PSPDFKit. You’ve probably never heard of it, but you’ve almost certainly used it. Ever opened a PDF inside Dropbox? That was his technology. Viewed a document on a Lufthansa cockpit iPad? Also him. The company’s tools ended up on over a billion devices. He raised $116 million from Insight Partners, stepped back, and promptly burned out.
He disappeared from tech for three years. Just... gone.
When he came back in 2024, he dove headfirst into AI. And late last year, he built something as a weekend project that has since become the fastest-growing repository in GitHub history. He created a personal AI assistant that runs on your own computer, connects to your messaging apps, and actually does things for you. Not just answers questions. Does things. Sends emails. Books reservations. Manages your calendar. Checks you into flights.
He originally called it Clawdbot, a playful nod to Anthropic’s Claude. (Specifically, it was inspired by the little lobster-claw monster you see when Claude Code is loading.) Anthropic’s lawyers didn’t find it quite as playful. So it became Moltbot, keeping with the lobster theme, because apparently that’s a thing. Then a few days later, Steinberger decided he just liked OpenClaw better. Three names in about a week. Welcome to 2026.
The numbers tell you everything. Over 150,000 GitHub stars, which makes it the fastest-growing repository in GitHub history. Adoption from Silicon Valley to Beijing. And just this past weekend, Steinberger announced he’s joining OpenAI. Sam Altman called him a genius and said the technology will become core to OpenAI’s products. OpenClaw will continue as an open-source project under a foundation.
From burned-out founder hiding from tech to getting recruited by Sam Altman. In about three months. I don’t think any of us fully appreciate what moment we’re living through.
So What Is OpenClaw, Actually?
OK so let me try to explain this in the simplest way I can, because I’ve already had about a dozen people ask me “what even is this thing?”
You know how ChatGPT and Claude are really smart, but they basically just talk to you? You ask a question, they answer. You have a conversation, maybe a great one. But then you close the browser and nothing happens. They don’t go do anything on your behalf.
OpenClaw is different. It’s an AI agent that runs on your own computer, 24/7, and you interact with it through messaging apps you already use: WhatsApp, Telegram, Slack, Signal. But here’s the thing that changes everything: you give it actual access to your email, your calendar, your files, your browser. And it acts on them. On its own.
Think of it this way. ChatGPT is like having a really smart advisor sitting across from you at a coffee shop. OpenClaw is like hiring a junior employee who shows up early, works all day, never forgets anything, and can juggle six projects at once. Except this employee costs about $20 to $200 a month in API fees and never asks for a raise.
It connects to a large language model like Claude, GPT, or Gemini to do the thinking. But unlike those chat interfaces, OpenClaw has hands. It can execute commands, manage files, send messages, and chain together multi-step tasks without you hovering over it.
Steinberger describes it as “AI that actually does things.” After a week of running my own instance, I’d say that’s about right. With some very important caveats I’ll get to.
From Costco to Command Center: How I Set Up My Own OpenClaw
I’d been tracking OpenClaw since it first started blowing up. And like apparently half the tech world, my first move was to go buy a Mac Mini.
Two weeks ago, I walked into my local Costco and grabbed one off the shelf. I’m not kidding. If you’ve been following this story, you know there’s been a genuine run on these machines. The M4 Mac Mini has basically become the official hardware of the OpenClaw movement because it’s small, sips power, runs 24/7 without complaining, and sits on your desk like a five-inch-square brain. Some higher-memory configurations are now backordered two to three weeks.
Mine sat on my desk for about a week while I finished client work. (I kept looking at it like a Christmas present I wasn’t allowed to open yet.) But when I finally carved out a Saturday to dig in, I started where I always start with new tech. YouTube.
I watched a ton of setup videos, including an excellent walkthrough from Matthew Berman that really helped me understand the architecture. Then I sat down with Claude and started mapping out what I actually wanted to build.
Legal research for client work? Yes. General AI and industry research? Yes. Daily personal and work email briefings with calendar integration? Yes. A custom CRM that tracks every potential client interaction? Yes. An agent that helps draft content in my voice? Absolutely yes. Making the whole thing secure enough that I could sleep at night? Non-negotiable.
Here’s the thing about OpenClaw that nobody tells you until you’re knee-deep in it: this is not for the faint of heart. And I say that as someone who’s comfortable around this stuff. I’d read plenty of horror stories before I started. People misconfiguring things and getting credentials stolen. API bills spiraling because an agent decided to check the time 400 times overnight (that actually happened to someone). One guy gave it access to iMessage and it blasted over 500 messages to random contacts, including his wife. That’s... not great. Gartner flat-out called it an “unacceptable cybersecurity risk” and recommended businesses block it immediately.
So I wanted to start small and build up deliberately.
How I Locked It Down
I basically treated this like onboarding a new employee who needs supervised access to sensitive systems. Maybe that sounds like overkill. It’s not.
I bought a completely separate machine and created fresh accounts for everything. No shared credentials with my personal or work systems. None. My OpenClaw instance got its own Gmail workspace account that can connect to my other accounts, but only in read-only mode. It can see my email. It cannot send from my real accounts. I gave it its own Claude subscription too, separate from my personal one. Then I started adding capabilities one at a time, testing each one, running security checks after every addition.
For communication, I set up a dedicated Telegram channel and gave the bot strict instructions: only accept commands from me through Telegram or the web interface. Never via email. That one matters more than you might think, because one of the biggest attack vectors with these agents is something called prompt injection. Basically, someone embeds hidden instructions in an email or document, and the agent interprets them as a command from you. It’s sneaky and it’s real.
I also had the bot create and manage its own passwords for connecting to services. I added TeamViewer (the free version) and Tailscale for remote access when I’m not at my desk. And honestly, I’m adding new security layers almost every day as I learn more.
Is this paranoid? Probably. But when you’re handing an AI agent the keys to your professional life, paranoia is just good hygiene.
Meet the Team
OK, here’s the fun part. What am I actually building with this thing?
My focus has been personal automation. Stuff that runs in the background while I’m doing client work, giving presentations, or writing. Think of it as building a small team of specialists who each own one job. I named them all, because of course I did.
Scout is my research agent. Every day, Scout wakes up and scours the internet for news about legal AI, generative AI developments, law firm technology, and California bar ethics opinions. But it doesn’t just collect links. It reads articles, scores them for relevance, and stores the good ones in a central database. By the time I sit down with my coffee, Scout has already done an hour of work I used to do manually. That alone has been worth the entire setup.
Quill drafts content. When Scout surfaces something interesting, Quill turns it into a first pass for my newsletter, a LinkedIn post, or a tweet. I’ve trained it on my writing style (which means no “synergy” and absolutely no “leveraging paradigms”), and if I ask, it can generate header images to go with an article. I still edit everything. But starting from a solid draft instead of a blank page? That’s a real time saver.
Ledger is my CRM, and honestly this one might be the most useful of the bunch. It tracks every contact, every interaction, every pipeline opportunity. It knows which firms I’ve talked to, what we discussed, where they are in the buying process, and when I should follow up. It’s relationship management that actually remembers. Which is more than I can say for most CRM software I’ve paid good money for over the years.
Sentinel handles competitive intelligence. What are competitors doing? New services, pricing changes, content strategies, speaking engagements. I don’t like being caught off guard, and Sentinel’s job is to make sure I’m not.
Archivist manages my knowledge base. Every research finding, every draft, every insight gets indexed with vector embeddings (that’s a fancy way of saying it can search by meaning, not just keywords) so I can find anything later. It’s like having a personal search engine for everything I’ve ever read or written.
Briefing is my morning assistant and might be my favorite. Every day it pulls together what Scout found overnight, what Ledger noticed in my pipeline, what’s in my email, what’s on my calendar, and sends me one clean digest. I read it with my coffee. I genuinely look forward to it now.
There are more agents, but you get the idea.
Are these perfect? Oh, not even close. I’m tuning them constantly. Some days Scout surfaces completely irrelevant articles. Quill occasionally produces drafts that sound way too polished (I have to remind it to sound more like me and less like a McKinsey deck). Ledger sometimes logs the same interaction twice. I spend time every single day adjusting, tweaking, fixing things.
But here’s what I keep coming back to: I am getting meaningfully better at this every day. Give me another week or two and this system will be light years from where it is now. That’s the nature of this stuff. It compounds.
Should Your Law Firm Be Using This? No.
I’ll be direct. If you run a law firm, you should not be deploying OpenClaw right now. Full stop.
I know that’s a weird thing to say in an article where I’m clearly excited about it. But part of my job, the part I take most seriously, is telling you the truth. Not just the parts that sound good.
The security concerns are real. OpenClaw requires broad permissions to function, and misconfigured instances can expose API keys, credentials, and access to sensitive systems. Cisco’s threat research team found a third-party OpenClaw “skill” that was quietly stealing data without users knowing. Researchers found over 340 malicious skills uploaded to OpenClaw’s marketplace. And one of OpenClaw’s own maintainers posted a warning on Discord that if you can’t understand how to run a command line, the project is far too dangerous for you.
For a firm handling privileged client communications, that should stop you cold. Your ethical obligations around client confidentiality under the ABA Model Rules aren’t suggestions. They’re requirements. And the legal liability picture for autonomous AI agents? Completely undeveloped. If your agent sends the wrong email or modifies the wrong document, that’s a malpractice question that literally nobody has answered yet.
Then there’s the leadership question. Steinberger just joined OpenAI, and the project is being handed off to an open-source foundation. Long-term, that could be great. Short-term, it means a platform that’s barely a few months old just lost the person who built it. That should give anyone pause.
Here’s what I keep telling people: there’s a massive difference between someone like me, a technically proficient AI consultant, experimenting with this on a dedicated, sandboxed machine, and a 15-person family law firm deploying it to manage client communications. Massive.
But Here’s Why You Should Be Paying Attention
OpenClaw is a preview of where all of this is going. The idea of autonomous agents that can run multi-step workflows, talk to your actual business systems, and operate without you babysitting them? That’s coming whether firms are ready or not.
Think about what I just described with my own setup. Research that happens while I sleep. CRM updates that happen without me touching anything. Morning briefings pulled from five different sources. Content drafts in my voice. Every single one of those use cases will eventually ship in enterprise-grade, secure, compliant packages. I’d say we’re 12 to 18 months away from that.
The smart move right now isn’t to deploy OpenClaw at your firm. It’s to understand what it does, watch how it matures, and start identifying which workflows in your practice would benefit most from this kind of automation. Because when the enterprise-ready versions show up, and they will, the firms that have already thought through their use cases will move fast. The rest will be playing catch-up.
What to Do This Week
Go watch a demo of OpenClaw to see what autonomous agents actually look like in practice. Matthew Berman’s YouTube walkthrough is the one I’d start with.
Pick three to five repetitive tasks in your practice that eat up associate or staff time: research gathering, status updates, follow-up reminders, intake processing. Write them down. These are your future agent candidates.
Ask your IT team (or your security advisor, or both) what your firm’s current policy is on autonomous AI tools. If you don’t have one, start writing it now. Shadow deployments are already happening at firms whether leadership knows it or not.
Put a quarterly check-in on your calendar to track what’s happening with enterprise agent platforms from OpenAI, Anthropic, Google, and Microsoft. This space is going to move very fast.
We’re Living in the Future
I’m writing this article right now while Scout is scanning the web for tomorrow’s research. Ledger just logged a new contact from an email that came in 20 minutes ago. Briefing is already pulling together what I’ll read with my coffee tomorrow morning.
A month ago, none of this existed in my world. Now it runs quietly in the background on a $600 machine the size of a paperback book sitting on my desk.
We are at the very beginning of something big. OpenClaw, whatever it becomes under OpenAI, is a preview of what’s ahead: smart, specialized agents that will work alongside solo practitioners and large firms alike. The tech isn’t ready for most organizations today. But the direction? Unmistakable.
I wanted to share my early experience because this is one of the most exciting and genuinely interesting things I’ve seen in years of working in this space. I’ll keep building. I’ll keep testing. And I’ll keep telling you what I find.
The future showed up. It runs on a Mac Mini. And it’s just getting started.
Why I write these articles:
In this article, we looked at the gap between what autonomous AI agents can do right now and what’s actually safe to deploy at a law firm. OpenClaw proves the concept works. My agents research, draft, manage contacts, and brief me every morning on a machine that costs less than a decent office chair. But the security model, the liability questions, and the ethical obligations around client confidentiality mean this is a “watch and prepare” moment, not a “deploy on Monday” moment.
The market is noisy, but the path forward is usually simpler than the hype suggests.
If you want help sorting this out:
Reply to this or email me at steve@intelligencebyintent.com. Tell me what’s slowing your team down and where work is getting stuck. I’ll tell you what I’d test first, which part of the agent and AI stack fits your workflows, and whether it makes sense for us to go further than that first conversation.
Not ready to talk yet?
Subscribe to my daily newsletter at smithstephen.com. I publish short, practical takes on AI for business leaders who need signal, not noise.