OpenClaw 5.12 Update: What You Need To Know…

OpenClaw 5.12 just dropped and I want to be real with you about where things stand here. It's been a rough stretch. The last several updates have been buggy. People have been frustrated.

Setups breaking, memory getting wiped, channels disconnecting. The community has been pretty vocal about it and now Hermes agent from News Research has overtaken OpenClaw on OpenRouter. More API calls are now going through Hermes than OpenClaw right now. That's a first.

So 5.12 matters because this is OpenClaw's answer. Peter Steinberger and the team called it faster, calmer and harder to wedge. And after digging through the change log, I think they're focused on the right things. Let me show you what change you can make your mind up.

The biggest thing you'll notice here is that OpenClaw just got way lighter according to this change, right? Before this update, when you installed OpenClaw, it pulled in everything. So WhatsApp libraries, Slack libraries, Amazon Bedrock, Anthropic Vertex, all of it. Even if you never even used those channels.

So your install was bloated with stuff you didn't need. It made install slow, it made startup slow and it created more chances for things to break. OpenClaw 5.12 rips all of that out of the core. Now when you install OpenClaw, you still get the core.

That's it. WhatsApp gets installed when you set up WhatsApp. Slack gets installed when you set up Slack. Bedrock gets installed when you connect to Bedrock.

So each piece only shows up when you actually need it, which is kind of what you hoped for anyway. For you, that would mean a faster install, faster startup, less memory use and fewer random crashes caused by libraries you weren't even using. If you've been using OpenClaw on a small server, a Raspberry Pi for example, or an older machine, this would be a genuine improvement as long as it works in practice. Sometimes there's a big difference between what you see inside the changelog and what actually happens in reality.

So I'll come on to that later. Telegram got a big reliability fix too and this one is important because Telegram has been one of the most frustrating channels lately and here's what's happening. So your agent runs on something called a gateway. That gateway has a lot of jobs.

It handles your messages, it runs your tools, it manages your memory and one of those jobs is checking Telegram for new messages. The problem was, if the gateway gets busy doing something else, it would stop checking Telegram. Messages would come in and just sit there. Your agent would not see them.

People would text your agent and get nothing back. OpenClaw 5.12 moves the Telegram message checking into its own separate worker. It runs on its own. If the main system gets busy, the Telegram worker keeps going.

Messages get saved to a local backup file as they arrive. So even if the main system hiccups, your messages are safe and waiting. On top of that, the store detection got smarter. Before, if your agent sent a message out, that activity would trick the system into thinking everything was fine, even if incoming messages were stuck.

Now, it only looks at whether new messages are actually coming in. If they're not, it knows something is wrong and tries to fix it. And there's a fix for group chats. If your agent was in a Telegram group with require mention turned on, meaning it should only respond when someone tags it, it was still trying to download images from messages that weren't meant for it and failing, sometimes sending error messages to the group, which is pretty embarrassing.

That's fixed according to this update. Now, here's something that matters for anyone whose agent has been stalling mid-conversation. You know that thing where your agent just stops, it's thinking, the little dots are spinning, and then nothing happens. No reply, no error, just silence.

OpenWorld 5.12 adds stalled stream recovery. So if the AI model your agent talking to stops sending data, maybe, for example, the connection dropped, maybe the server hiccuped, your agent now detects that. And instead of sitting there forever, it tries a different approach. It rotates to a backup model or a backup login.

And if that doesn't work, it shows you an actual error message instead of just going quiet. One community member, Tashin, said this is the fix that matters most for anyone running agents in production, because an agent that suddenly hangs is an agent you can't trust. And he's right. Now, if you want help setting up OpenClaw, running it for your business, getting more leads, getting more customers, setting up workflows for Hermes, Agent, and OpenClaw, that's what we do inside the AI Profit Boarding.

Four coaching calls every week. Right now, a lot of the conversations are about exactly this sort of stuff. How to keep agents stable, how to actually use them in real use cases, how to configure it properly. We've got daily tutorials in there, 3,000 business owners in there building with AI agents, and a member map so you can find other OpenClaw and AI automation users near you.

Link in the comments description or go to the AIProfitBoarding.com to get access. The OpenAI and Codex setup changed. So, when you set up OpenAI in OpenClaw now, it defaults to the chat GPT subscription login, which is the Codex route. Before, we would ask you which method you actually wanted, but now it assumes you're on a chat GPT subscription and starts there.

If you want the API key route instead, you can still choose it, but the default is now the subscription path. This is smart because most people using OpenAI through OpenClaw are on a chat GPT subscription. They're not buying API credits separately, so the default should match what most people actually have. The web chat dashboard got a scroll fix that sounds small but was driving people crazy, so when your agent is typing a long reply and the text is coming in, the chat would auto scroll to follow it.

But sometimes you'd want to scroll up to read something earlier and the chat would keep yanking you back down to the bottom. You couldn't read what you wanted. Now, you can choose, right? You can stay at the bottom and follow the inputs as they come along, or you can set up always follow new output, or you can turn auto scroll off completely and use a new messages button when you're ready to jump down, which is pretty simple.

But it matters, you know, when you're using this every day. Some people are using this for hours every day, right? The command line also got smarter about errors, so when something goes wrong now, your settings, for example, are broken, a channel won't connect, an add-on fails, OpenClaw should tell you what happened and gives you the exact command to fix it before you'd get a confusing error and have to figure it out yourself. Now it points you to the answer.

There's a security pass in the update too, so Windows home folders are now blocked from being shared with your agent sandbox. That means your agent can't accidentally access your passwords, your SSH keys, or other sensitive files stored in your Windows user profile. Provider credentials now go through a structured system instead of just reading random environment variables, and the pairing process where you connect your phone or browser to your gateway got stricter. Plug-in installs got more reliable too.

They now support pmpm11, which is the package manager OpenClaw uses internally. Add-on dependencies handle better, and the code safety scanner that checks add-ons when you install them no longer blocks legitimate add-ons because of code that lives deep in their dependency chain. Now let me talk about what the community is saying here too, because the reaction to the update tells a story. Andre, for example, on X, and I'm paraphrasing here, says, has anyone tested this and can confirm it didn't break their setup?

And that question tells you everything about where trust with OpenClaw stands right now. People are afraid to update, right? And a lot of people are spending more time fixing these updates than actually using OpenClaw, and I think that's been a common theme over the last few updates. So before you think about updating, well, I'll come on to that in a minute.

Felipe said he's never had a smooth update since the first version, and now Slack is broken for him. That was another community reaction on X to the update post. So you have to be realistic here and understand that OpenClaw's own framing of faster, calmer, harder to wedge shows they know stability is the priority right now. And let's address the elephant in the room here.

Hermes agent just passed OpenClaw on OpenRouter. More API calls are now going through Hermes and OpenClaw, and the reason is simple. It's bugs. When your tool breaks your setup every other week or every other day for some people, people will switch, right?

They don't switch because the other tool has better features. They switch because they're tired of fixing things. They're exhausted, right? And Hermes has been way more stable.

They ship usually about once a week, and they tend to break less, and I think they learned from watching OpenClaw's mistakes. Now, does this mean that OpenClaw is done? No. OpenClaw still has the broadest channel support of any agent tool.

It's got the biggest community, over 100 contributors in this release alone, and it's still one of the biggest in open source AI. But the thing is, features do not matter if people cannot trust the updates, and that's a core issue here. OpenClaw needs a streak of clean, stable releases, not new features, not flashy announcements, just updates that actually work and don't break things. Now, 5.12 seems like a step in the right direction.

The lighter installs, the telegram isolation, the store recovery, these are stability improvements, the right kind of work. Whether it's enough to win people back from Hermes depends on the next few releases. If the next three or four updates are clean, trust rebuilds. If the next update breaks telegram again, what more people are going to leave?

And here's what I tell anyone watching this. Number one, if your setup works, don't update yet. Wait a few days, watch what people report. Andre, for example, asked the right question.

Let other people test it first, right? Number two, before any update run, OpenClaw backup create. Every time, no exceptions, right? Because you always want to have a backup of what's working for you right now.

And also, I would make sure that you save that, that you even write down, for example, which version you want. So that brings me on to number three, which is know your rollback version. Whatever version is stable for you right now, write it down. And number four, if you do update and Slack, for example, breaks, check OpenClaw Discord for the fixes before you try anything else.

Chances are there's probably someone else who's already found the solution and posted about it. And five, I wouldn't chase tools. Chase stability. You know, one of the things that I've learned from this whole situation is like, you've got OpenClaw, you've got Hermes, you've got Claw Code, you've got something else, right?

But the people who win are the ones who really like master a tool, learn it deeply and build real workflows. You've probably seen my other videos on Hermes recently, for example. That's skill compounds, right? If you jump ship every time something breaks, that means you never get deep enough to be, you know, really capable with these AI agents.

Now, for me personally, I have a lot of time to test this stuff, so I'll test it for now. But for anyone watching, I'd be like, right, let's focus on this. Let's only do that. If I had to pick between all the tools right now, if you want the most stable, I'd go with Claw Code, right?

Super powerful agent, nothing wrong with that. If you want something that's open source, gives you freedom for customization, I would go with Hermes. And I wouldn't use OpenClaw until you see like reports, genuine reports, where the consensus is OpenClaw is much more stable. And the people who figure out agents now, while things are still messy, are the ones who will have a massive head start, right?

Like every update is a lesson, every fix is a chance to understand your setup better. And when the tools stabilize, and they will, you will already know how to build, troubleshoot, and automate. Now, if you want to learn with real support, come join us in the AI Profit Boardroom. Four coaching calls every week.

This week, we're covering exactly what we talked about, how to update, you know, how to set up AI agents, how to use this in the real world, how to generate customers and leads with this stuff, how to keep your business running with these sort of AI agents. And we've got a 30-day OpenClaw roadmap. We've got really useful guides for the Hermes agent. Daily new tutorials dropping all the time.

3,000 members in there. I'm apt to find people near you doing the same sort of thing. Link in the comments description to go to theaiprofitboardroom.com to get access. And again, before you update, run OpenClaw backup.

You can even, for example, just set it up on a separate server or, for example, a separate user profile, see if it works, and then go back to it. I'll see you in the next one. Cheers. Bye-bye.