How to Build Custom OpenClaw Dashboards: Cloning Clearmud OS, Pulse OS, and a Thumbnail Plugin

Good morning, good afternoon, and good evening. This one is a companion post to Monday's live stream, where I brought it back to OpenClaw. Over the weekend I upgraded my OG system to the latest version, discovered the new built-in dashboard plugins, and got excited enough to spend two hours building in front of chat. Here is what I did, what worked, what broke, and where I am going with it.

The upgrade I was dreading just worked

I had been sitting on OpenClaw 7.2, or 7 something, for a long time. It was one of the more stable versions and it just worked with how I had everything set up. I do not chase every update. Early on I found that every time I chased a new release, something would happen more often than not, so I stopped updating frequently unless a newer model required it.

This weekend I made the commitment to move to the latest version. I was expecting things to break. I was expecting to spend at least five hours getting my old systems working on the newer build.

Instead, I triggered the update with Muddy, my COO agent, walked away, and went and grabbed lunch because I did not want to watch it. I came back and it was just done. No errors, no issues, no follow-up. Muddy and team were all there working. All 44 active automations carried over.

That is the whole story of the migration. Boring in the best way.

The idea: bring Clearmud OS into the OpenClaw gateway

The reason I built Muddy OS in the first place, back in January and February, was that I did not like the OpenClaw dashboard. I could barely customize it. Now OpenClaw lets you build your own dashboards, create tabs, and pin them inside the gateway. That changes the math.

Clearmud OS always had one pain point I never took the time to vibe code: I could not chat with Muddy inside it. Muddy OS was built to help manage the system, so chatting there was always possible. Clearmud OS never had that.

So I started small. I used Elon, my CTO agent, as the test account and began rebuilding each Clearmud OS tab as its own dashboard. Then I thought, wait a second, let me just see if I can clone Clearmud OS one to one as a single dashboard inside my OpenClaw gateway.

It worked. I did not think it was going to. I thought we were going to be limited, but apparently we are not.

Now I can pop open a chat panel next to the content calendar and the rest of the tabs and talk to Muddy and team as I work through it. I can hide the chat when I do not need it. That is exactly the thing I always wanted.

What is not working yet

I want to be transparent about this because I am not relying on it yet. This is testing and tinkering.

• The content calendar is not working.
• The dashboard overlaps the gateway's buttons up top, so I need to add some spacing.
• There are still weird little bugs and glitches to fix.

The front end looks like a one to one clone, to be honest. The real problem is the data layer. A while back I got away from Clearmud OS being a bucket of JSON files and moved the content calendar and other apps to SQLite. So the next step is figuring out how a dashboard plugin ties into a local SQLite server. Front end solved, storage strategy is the next conversation.

Building Pulse OS live on the local demo instance

I keep a bare local install of OpenClaw, somewhere around 9.1 or 9.2, that I use for demonstrations. I do not rely on it, which makes it the perfect sandbox. On this stream, the agent on that instance was running ChatGPT 6 Astra, and I used it to build two dashboard plugins from scratch.

The first was Pulse OS. I dictated the brief using OpenClaw's built-in dictate feature instead of my usual Whisper Flow, and the idea was scattered on purpose, from a beginner's perspective:

• Have agents or sub agents power the dashboard on a frequency we define later.
• Invoke the last 30 days skill to see what is being discussed, ideally including X.
• Cross reference that with Google Trends, toggled to YouTube search instead of Google search, targeting roughly the last one to three months.
• Check what shows up on a blank YouTube homepage.
• Run all three inputs through a pipeline that produces video ideas, after digesting the Clearmud channel so it learns the kinds of videos I want to make.

The goal is simple. I want to open one place every morning and get a quick TLDR of what people are talking about, without opening social media and without manually prompting and checking my timeline. Finger on the pulse with as little human involvement from my end as possible.

The mic missed a few beats, the agent named it Pulse-OS with a hyphen, and the first version came back looking more like a plan doc than a web app. But I was not mad at it as a first go.

The voice mode surprise

While building, a voice interface woke up and started talking back to me. I had not set up an OpenAI API key on this desktop version, so it must be pulling from Codex OAuth or OpenAI OAuth. I have no clue exactly how it is wired. I am happy it works, and I am hoping I do not get a surprise bill.

The voice mode also showed me the limits of talking to an agent. I asked it to lock all agent and sub agent browsing inside OpenClaw's embedded browser instead of popping external windows. It kept explaining what it would do, kept hitting timeouts, and kept telling me it could not verify anything. At one point I just said: talk less, take more action, just do what I am asking. It could not get there. The browser lockdown is still not done.

Lesson: it is nice to have, but I am not ready to drive a build by voice alone.

A thumbnail generator plugin

The second plugin was a thumbnail generator. The brief was: I type in a video title, and you create thumbnails based on your own interpretation of what makes a successful YouTube video, pulling elements from our channel so I do not have to upload anything.

The first result came back as a form that spins up an agent to do the work, and the agent delivers the image in chat and refreshes the preview. Interesting model. Not mad at it. The thumbnails themselves were a mixed bag. It applied the rule of three, hook text plus a glassmorphic desktop UI plus me, but none of it was relevant to the actual video. When I asked for a thumbnail on ChatGPT 6 Astra building OpenClaw dashboards, I expected it to pull the OpenClaw logo and the GPT logo. It did not. And the face it gave me was rough, probably because of one weird thumbnail reference it grabbed.

So I gave it more direction:

• Grab our highest performing thumbnails from the past six months and use them as reference points.
• Create three versions of every request for A/B/C testing.
• Lay it out like a Google search bar at the top, a filter toolbar in an accordion below it for composition and words to emphasize or avoid, a loading state, and then three vertical columns side by side with the results.

After the second pass it was already doing better. One idea, three directions. Boom.

How OpenClaw handles dashboards now

A few things I learned about the dashboard system itself:

• Dashboards open as pinned sessions with a dashboard attached, not as dedicated app pages. Chat opens on one side and the dashboard on the other. I actually like this better than tabbing between a chat and a full screen app.
• You can assign icons and colors to dashboards. The icon set is limited right now, and there has to be a way to add more, but even little tweaks like a pink icon and a code block make it pop.
• Google Trends links with saved filters carry over. I set worldwide, past 30 days, all categories, YouTube search, and opening the link in a different browser kept every filter.
• The desktop app is a direct one to one clone of the web dashboard. If I open the loopback address in my browser I see the same thing. That consistency is a big deal to me.
• There is a status view that gives you a finger on the pulse of what each agent is up to without clicking into every tab.

The gateway did freeze once mid stream. I resisted force quitting because I could still resize the window, which felt like a sign it was not fully bricked. It turned out there was an update waiting, and after a restart the work had actually finished in the background. Be patient before you force quit.

My design feedback on the first Pulse OS build

First impressions of Pulse OS: too much fluff, too much empty space. Fluff to me is when you have to move your eyes down three rows before you get to anything relevant. That is wasted space.

What I asked for instead:

• Tabs across the top like a web browser, not a vertical menu bar.
• Each tab gets straight to the point.
• If you build an editor's pick section, make it look like a New York Times homepage with multiple columns and rows.
• Each idea should look like a sticky note: a TLDR of the hook, what the video is about, title and thumbnail ideas, and source links for where the idea was discovered. Those source links are critical.
• The today section needs more detail, and the evidence should live with the sources.

The second pass looked a lot better. I still want bigger text and a quote block so the hook pops, but it is heading the right direction.

OpenClaw vs Hermes, and why I use both

Chat asked, and I will be honest about where I stand. I have zero automations scheduled on Hermes personally. I deploy Hermes agent systems for clients, with one agent and a few sub agents helping me manage those accounts. When I started that first contract, OpenClaw was not as stable as it is today, and for client work Hermes is just stable. I am a Dexter's Laboratory type of person. I do not mind fixing things on my own systems, but I do not want to fix things for a client.

My one gripe with Hermes is that the desktop app looks nothing like the web dashboard. OpenClaw's are one and the same, and for me personally, OpenClaw's dashboard wins hands down. The visual design has always been the thing.

But listen, I do not think there is a right or wrong answer here. Pick one you vibe with and roll with it. No harness kills another harness. Attention shifts. If all a creator does is chase that attention, I find it tough to trust their opinion. That is just contributing to AI slop.

And if you do not find a harness you vibe with, build your own. It has never been easier. Take the best bits of each harness and bring them together into something that is yours. Even Mr. Hermes himself in chat said he is contemplating forking OpenClaw, which was music to my ears.

A note on skills

Someone suggested installing a skill like caveman or unslop to stop agents from being wordy. The caveman skill is cool, and the site is clean, but I do not install third party skills. I have found that newer models degrade when they lean on skills that were written months ago. The only skills I keep are ones for repetitive workflows where I need the output to be consistently what I defined. For verbosity, I just tell my agents: be direct, to the point, as brief as possible. Less is more.

What I am letting go of

Here is the honest takeaway from two hours of building. I am loving OpenClaw more and more the more time I spend in it. I think it is time to let go of the dashboards I built in Muddy OS and Clearmud OS and recreate them natively in OpenClaw. The one to one clone was fun, and it proved the concept, but a one to one clone is really just me not willing to let go. My YouTube script pipeline is a good example. If it lives as an OpenClaw tab, the agent responds right there, the styling is better, and I can ask for output in a quote block with all the options.

Will I find a way to make everything work? Not sure yet. Next on the list is OAuth login through Google Workspace so I am not working out of the administrative account every day, then the SQLite connection for the Clearmud OS plugin.

If you have been sleeping on OpenClaw, now is the time to start dipping your toes. Not the dark side. The bright side.

Thanks for hanging out with me on a Monday. We will be back at it Wednesday, same time. Until then, happy building. Cheers.

Watch the full live stream: https://www.youtube.com/watch?v=gCXnMgtUjMc