Is Openclaw ready for Regular Business?

A couple of weeks ago, I watched Jensen Huang say that every business needs to be working on an "Openclaw Strategy". With all of the horror stories I heard in the early days of Openclaw (only a few months ago), I was leary to set up a normal Openclaw setup but thought that a Nemoclaw setup, Nvidia's own safeguarded version, might be more reasonable.

I happened to have a Hostinger VPS sitting empty so I decided to deploy Nemoclaw on it to isolate the blast radius. If you're not familiar with Nemoclaw, it is Openclaw, inside a sandbox, inside a Docker container, ruled by a PowerShell management system. It's basically putting Openclaw inside a kennel, inside a dog pound, with an armed guard standing outside. As you might guess, that makes installation and operation more cumbersome. In fact, it was so cumbersome that after two days of fumbling around and breaking it 7 times, I finally got it all set up. Once I did have it set up, I realized that it was so locked up, it actually wasn't that useful. If you wanted to use it to create an agent or agent system that was only tied to one application and use case, it would probably be a very safe way to do it. If you want to do more, it's a pain in the butt.

Once I discovered that and became a little more comfortable with the Openclaw system, I deleted Nemoclaw and installed Openclaw. A simple installation of Openclaw is fairly simple to do. Enter one line in the terminal of the host machine, go through the onboarding (make sure you have an API key ready) and it's running. You can make it more complicated if you want to put it inside Docker. However, it's not hard to set up that one agent and the Openclaw environment. However, when you start playing with the setup, it becomes more complicated. Most of your work will be in the terminal. You can control some things in the web UI now, and there gets to be more with every update. Regardless, most of your setup changes will be in the terminal.

I thought that I was being smart setting up a space in Perplexity with the Openclaw official docs site as the reference material, but for some reason, Perplexity wanted to make everything harder than it has to be. Same was true for using Claude to make changes. What I eventually learned is that the smartest thing you can do is just to read the official docs yourself. It will actually save time and frustration to just look up what you need to know. Like Nemoclaw, I broke my Openclaw almost daily by making config changes to try and get what I wanted. Eventually I figured out that if I'll just use the built-in configuration update walkthrough, I could make changes without breaking things and then fighting with it.

A Few Things I Learned

I started out using my ChatGPT account to set up the model provider through OpenAI Codex which lets you use GPT-5.4. GPT-5.4 is a very good coder and reasoning model. It's really capable at a lot of things. It's also really lazy. Using it to do everything is like having a teenager working for you. It would say "Sure, I'll get that done." Then it would get started, hit something hard, and quit. It wouldn't tell you it quit. Sometimes it would even say it finished. It didn't. It had good intentions and bad follow-through. In addition to that, when I was trying to use it for everything, I got rate limited after about 24 heavy use hours and it shut me off for 48 hours.

After listening to all of the "Openclaw experts", I added Anthropic to the mix. I started running Sonnet 4.6 (because I knew Opus would be expensive). I did this through OpenRouter so I could eventually have access to other models like Gemini 2.5 Flash for heartbeats. What I didn't realize is that when adding OpenRouter, it changed my defaults to openrouter/auto. The auto mode decided to use Opus 4.6 or Sonnet 4.6 for everything. That cost me about $70 in two days. When I figured that out, I shut off openrouter/auto.

What I'm actually running now is a three-layer system with a main agent, an orchestrator agent, then a bevy of subagents including a front desk agent, a coder agent, an n8n architect agent, a researcher agent, a deep researcher agent, a generalist agent, and an analyst agent. The main agent is running on GPT-5.4. The orchestrator and coder agent are running on GLM-5.1 (more capable than Opus at 1/20th the cost). The n8n agent is running on Gemma 4 31B. The research agent, front desk agent, and generalist are running on Gemma 4 26B. The analyst is running on Qwen 3 Coder. The heartbeat is running on Nvidia Nemotron 3 Super. All of them have a handful of other free models as fallbacks in case I get rate limited. This gives me a very low-cost and surprisingly capable system as opposed to the incredibly expensive setup I had before. So far, I like this and my orchestrator has been very good at completing tasks running GLM-5.1.

Something else I learned is that skills and the soul.md and agents.md files are important. I won't go into the full explanation of these, but I had to add lines to those documents on both my main agent and orchestrator that told them their main priority was to GSD. They don't quit unless they're stuck and need help from a human. They keep going until they completely finish a task. In addition, your orchestrator needs to know how to pass on errors, failure history, etc. to your coder agent, otherwise it will suck.

Keeping Things Safe

I currently have an n8n server set up on the same VPS. I use that to connect Openclaw to things like my email, calendar, OneDrive, Gmail, etc. That accomplishes two things. It keeps my API keys and other credentials out of the hands of my Openclaw. In addition, it takes a lot of the workload off of Openclaw which saves tokens. n8n can run on a timer, go grab all of my email, calendar info, and even my notes out of Notion, combine those into a JSON payload and send it to my Openclaw. Then Openclaw only has to process the payload and deliver it via Telegram.

Another safety precaution I've taken is to not give it a browser. Instead, I used Brave's API to pull in AI search results so they are already filtered before they make it to the agent. It's not 100% foolproof but it's a lot safer than turning an agent loose on the whole internet with its own browser.

Another good practice is setting up a workflow to back up your workspace to GitHub. I had to set up my Openclaw agent "Buddy" with his own GitHub account and his own repos. He is only allowed to pull and read my other repos due to org settings which keeps it from being able to do anything crazy to my live projects but also allows it to be helpful in resolving issues.

My Verdict So Far

To be honest, I think that Openclaw has huge potential and will be very helpful to me in setting up custom workflows inside my various businesses and I think it will continue to help me automate things (along with n8n). However, I don't think it's ready for completely non-technical users. If you're not comfortable in the terminal or don't understand how most of the things I've mentioned in this post work, you're probably going to get confused, frustrated, and quit. For the semi-technical nerds and the developers out there, this is a great solution. If you have a team of developers at your disposal, this is a very workable solution. If you just have a normal IT manager or IT team at your disposal, you're probably all going to get frustrated with it eventually. If you're a solopreneur with time on your hands and a desire to learn, you can make it work but expect to commit a couple of weeks to getting it working well. Everyone has to make their own decisions but I will say that the UI gets better with every update, the onboarding gets better with every update. Eventually, Openclaw or something like it will be more mainstream and ready for the everyday business user but I don't think it's there yet.