I moved my OpenClaw work to Grok Bot. It mostly worked Transcript, AI Summary & Key Points

00:00 I've run an OpenClaw agent every day since January, and I spent too much time babysitting it, fixing after updates, broken logins, silent failures. These are the worst. So five days ago, I moved that workflow to Grok Bot. Since then, across sixty-seven logged jobs, I needed to step in seventeen times, and I had to fix three things before I could start trusting it.

00:30 So what happened? Most of the work that OpenClaw was doing successfully moved to Grok Bot, and most of the babysitting didn't. If you're new, Grok Bot is xAI and Cursor's agents app, and bots are teammates with their own cloud computer, and they can sign in into your tools and do the work there. And how did my five days look like? Sixty-seven real jobs, forty finished without me, fourteen finished after I helped, five half-done, eight failed or got canceled, and fifty out of the sixty-seven didn't need my intervention

01:06 at all. Let's start with the smallest job that I gave it on purpose. I typed a simple prompt, "Find me a place to fix jewelry." And what actually went out to the research bot was a full brief. It added my neighborhood, how long the shop should have existed, reputation, what to avoid, and a ranking method. And those were my real preferences, near me, established, owner-run, no chains, no mall kiosks.

01:42 I didn't write any of that in the prompt. It came from context that the system already had about me. What came back was four ranked shops with opening hours, and I got a three-line answer and a recommended pick. And my time spent on this was zero minutes. The job itself is nothing. It's a simple task. Any agent can search a map. But what actually sold me is between the simple prompt and the actual search, the brief Every job in this video, including the real ones, went through the same translation of my simple prompt

02:22 and what the agent got as a full brief prompt. That brief was written by the setup that I built out after my first plan failed. My first plan on day one was that I tried to port my old OpenClaw agent wholesale, repo everything. Didn't work. It didn't translate one-to-one between OpenClaw and Grok Bot. So I rebuilt this as role-specific bots, and then that worked.

03:01 And this is how my setup looks like. There is Sharky, my COO, chief operating officer, and I talk mainly to him, and there are seven specialists that Sharky briefs. My lazy prompts come in, and real full context, thought-out briefs come out. And Sharky, the manager agent, writes the kind of structured brief a specialist needs better than I do. Agents know what agents need.

03:33 While we are looking at my team, these fun, happy-looking fellows, I think one aspect of software that is hugely underrated and underappreciated is overall pleasantness. So it's much more pleasant to look at these small little fellows or have animations like these than to work with OpenClaw. If I'm working with these guys every day, then that counts, and that compounds.

04:09 But coming back to my fancy-looking team here and what they actually do, all of them share one cloud computer, one browser, one system hosted on a data center IP, and this is important. I'll come back to it in a moment. Now, the three failures which I also encountered on day one. The first failure is the network. The cloud computer couldn't log in into my Twitter.

04:38 At first, I didn't know why. It didn't tell anything else other than, "We've temporarily limited your login." And no matter what I did, it couldn't log in. But then after some research, I found out that it is because the machine is in the cloud, uh, in a data center, so it shares a data center IP, and it usually gets blocked by many different services because it's more than likely that it's a bot using the service, not a human.

05:13 It's kind of funny that Grok Bot cannot log into Twitter with Twitter being the product of the same company, although developers said that they are working on a fix for that. The fix that worked for me was to install Tailscale on the cloud computer, and I also have Tailscale on my local machine, my Mac Studio, and I made my Mac Studio as the exit node.

05:45 So the traffic goes through Tailscale, through my local machine, my Mac Studio, which is always on, and all the services and all the websites that it connects to see my home IP and see the residential IP so they don't block my IP and my Grok Bot. Although the developers are working on a fix, but I'm sticking with Tailscale because even if Twitter is fixed, other sites will still see the data center IP, and I'll be solving crazy CAPTCHAs all the time, or even get blocked like with Twitter.

06:29 Second failure, inadvertent writing. A bot meant to read a tweet from Twitter, but it ran the write command instead. So a post went out from my account. But to be fair, the bot quickly realized its error and immediately deleted the tweet. And also on top of that, he had Tinkerer, my development-related bot, build a write guard, like a harness. So now nothing on this computer can post to Twitter unless I explicitly approve it.

07:07 And even the Grok Bot docs say to not use separate bots as a security boundary. Every login on the box belongs to every bot. It can be both good and bad. Third failure, the cloud computer And updates keeps your files that you have accumulated, but drops installed packages, or at least some of them. Happens twice to me. Tailscale was gone both times and Obsidian Sync was dead, and nothing tells you there is a problem.

07:45 My tinkerer ended up building a bootstrap script which it runs after every cloud computer update, and reinstalls everything that is missing, checks is everything is working, and it seems to be working fine. Theoretically, I could set up a routine, an automation or a cron job, depending on the name, to do that automatically after an update. But at least for now, there is no trigger after an update or after a restart to use it.

08:18 Also, on top of all this, on top of everything that the agents are doing on the computer, there sits an auto review, kind of like a reviewer of every command that they are running. It blocked about ten of my sixty-seven jobs. Full script runs, a gist upload, printing a private key. It was annoying, but it was correct, and it even caught my own tinkerer bot assembling a script from Hex to get around the auto review and the reviewer won, which is good.

08:53 From what I saw so far, this is the security level that I want. It blocks the agents when they are trying to be too smart, smarter than they need to be, and blocking that while still making sure that the task is done. Another nice thing buried in settings, and many people probably don't even realize it, is how easy it is to add more than one account per service.

09:24 You just go to Plugins, you select which connector you want to configure, and then you can add as many accounts as you want. I have two for Gmail, two for Calendar. Another win and a huge life simplifier is how easy it is to connect new services. For example, here, the only thing that I did, I posted a link to a documentation for DataForSEO for their MCP connector.

09:56 That's it. The only thing that I needed to do, I needed to enter the credentials into this small window that popped up, and after that, everything was configured and working. I didn't have to do anything else. Another win was how easy it was to give all the knowledge about me that, for example, OpenClaw knows about me right now, to give it to Grok Bot I keep everything in Obsidian and it now has almost 4,000 markdown pages about everything about me, my work, my personal projects, my life, and everything else And it

10:36 synced that easily into the cloud computer, and now it has full context on me and all my projects. And that's why the jewelry brief knew my neighborhood, and that's why it feels on day one like I'm continuing a conversation with an agent that knows everything about me, rather than starting from scratch and teaching the agent what to do, how to do it, what to avoid doing.

11:04 Third win that I noticed was research, and it's probably more about harness or maybe system prompt. But what I noticed with many other agents is that when an agent cannot do something, they will just skip that part or do something else or do something partially or do something without access to some critical part of the process because they don't have a login, for example.

11:36 With Grok Bot, what I noticed a few times, that when a keyword API broke, like, mid-job, it said that it's blocked. It didn't invent the numbers, it didn't skip the process, and it gave me a window to fix it, and then it finished after I fixed the key. Might be a coincidence, might be how harness is configured, but it was nice to see that realization that something is missing and letting me know that it needs to be fixed.

12:08 These are some of the groups of jobs that I give the agents. For content pipeline, it's something that I had on OpenClaw. I'm automating an Instagram posting for a page, not my personal one, and I'm just looking what's possible, what's being blocked, because there is no official API. So doing a lot of things manually, seeing how far it can go, approvals, keeping the session live and everything else.

12:36 So far, it works well. And the rest is automations, cron jobs, and if I want to monitor anything, like is my parcel coming? If it is, let me know. Find the email, find Slack, and let me know when it's coming. Now usage, since everybody wants to know how much it can do within the subscription limits. My weekly bar after all sixty-seven jobs was twenty-nine percent.

13:08 So with quite heavy lifting and setting everything up and configuring and bringing Obsidian in and setting up the agents, I'm up to thirty percent of my weekly usage. There is a seven-day trial of Cursor Ultra that you can use to get Grok Bot. But it lasted me about two hours and I needed to upgrade. So it's good to get your feet wet and to look around what it can do, but not for much else.

13:38 Money. Would I pay $200 or $300 a month for Grok Bot alone? No, I wouldn't. But I didn't add a new $200 to my monthly bill. I'm replacing my $200 Claude Max subscription with Cursor Ultra subscription because I wasn't using all my Claude Max Monthly limits anyway. I mostly use Codex, so I'm not getting rid of anything. I'm not blocking myself from using Anthropic models.

14:17 I still have that in Cursor Ultra, but now I also have Grok Bot included in those two hundred bucks per month. If for you it would be an additional two hundred, my recommendation would be different. Then use the trial on the exact workflows you care about before paying. What OpenClaw has that Grok Bot doesn't? A few things. Agent separation, which like I said, it could be bad, it could be good.

14:48 If all agents share the same machine, same logins, same memory, it's easier when you add a new one, but also you get less specialized agents, where if you set up every agent separately, you can give access to very specific memories, very specific code bases. It's a little bit better at separating all of that, but it comes at a price. Which funnily enough, price is the second big things that OpenClaw has for it.

15:22 It's free. You only pay for the model for whatever you use. It could be your ChatGPT subscription, it could be even your local model, and you don't pay for your monthly subscription for running a bot. And the third is those different models is also a big difference for now, that you can choose any model you want. It can be GPT Sol, it can be Kimi K3, it could be anything, could be even Grok.

15:53 With Grok Bot, you only get Grok 4.6, which is great for everyday work, but I still wouldn't put it at the level of GPT Sol Max, for example, or Ultra. If you are starting out today with Grok Bot, I have a few helpful rules for you. One, make a chief of staff, COO, CEO, gateway agent, whatever you call it, first, and then let it write the briefs. Let it create your sub-agents, because it will also give them good descriptions, and it's better when your chief of staff creates that than you give the descriptions.

16:37 Again Agents know agents better than humans. Two, before a bot gets a login or when you configure MCP or when you configure, like you saw Gmail or Calendar, make sure that it cannot post or reply or do anything that you don't want to do, something destructive or something that you cannot undo easily, and these should require your permission. It should never be able to do by itself.

17:08 And three, at least for now, I assume that they will fix it soon or automate it soon, everything stays and should stay in the workspace directory, 'cause that is not wiped out after a computer restart, and write a bootstrap script that recreates everything and reinstalls everything after the restart. It doesn't happen often, and if you don't install any crazy libraries, you might not even need that.

17:40 But be aware that this is something that you might need, and it fails silently, so better to know beforehand. So here's where I am after five days. I've moved all the workflows from OpenClaw to Grok Bot, and for now, I'm not shutting down OpenClaw. I'm running both in parallel, and if after a few weeks Grok Bot does the same job or more or better and still requires less babysitting, I'll switch to Grok Bot entirely.

18:16 For now, it feels like a fraction of repairs and babysitting of what seven months of my OpenClaw setup trained me to expect. So, so far so good. So my verdict after 67 jobs, Grok Bot is good enough to own research, monitoring, file-based workflows. It's great for all those everyday workflows, automations, admin stuff. It's not great yet for complex coding, but good with quick fixes.

18:54 It's useful enough now to move real work to. Grok 4.7, when it comes out, which is promised in a few weeks, might take Grok Bot to a new level, so my recommendation could be different in a month. Ask me then. For now, go assemble your nice-looking smart bots, and have fun bossing your team around