OpenClaw 3.11 Update: What's New (Release Notes Breakdown)

0:00Open Claw just released their latest update from 3.8 to 3.11. I'm going to be covering all the details in this video and what I'm doing to take advantage of this latest update. There are some serious stuff in this release and I'm going to be covering it all, including a security fix that if you're running this on any kind of server, you may need to update this right now because we're talking about a vulnerability that could give random people admin access to your agents. So, I literally just finished updating my Mac Mini setup. I had to run the Open Claw Docker fix because of some of the breaking changes and I was up late debugging the cron issues with my agent team. So, this timing is perfect. Tell me if you've been there before. So, let me walk through what's actually important in this release. The first thing is that security fix. This is critical if you're running Open Claw on

0:46a VPS or any kind of exposed setup. Cross-site Cross-site web socket hijacking and trusted proxy mode. Basically, untrusted origins could potentially get operator admin access to your system, which is pretty scary. Now, I run everything locally on my Mac Mini. I don't expose anything to the internet, but I know people in my community are running this on servers and on VPS setups. So, trying to access it remotely will happen. If that's you, you need to stop what you're doing and update this right now to add that extra protection of security. This isn't something you want to wait on. That vulnerability means someone could potentially hijack your web socket connections and get into the admin privileges. So, that's your entire agent system, your API keys, pretty much everything. So, I'd probably

1:32update that first if you're hosting it on a server. So, you can watch the rest of this video later, but I would stop right here and go do that right now. Now, for the features. The big one that's going to matter a lot for a lot of people is the first-class Ollama support. They've completely redone the onboarding flow with local and cloud plus models and I don't personally use Ollama, but I know a few people in our group do and I'm just running with the Anthropic OAuth token through my subscription and I use Opus 4 6 for all the execution and coordination. That's my main model. So, and that's what I was telling people on the boot camp call today. If you want to have a good open cloud experience, start with the best frontier model first and then work to optimize your token usage for other smaller jobs. So, use Opus as your main

2:18workhorse and then use a cheaper model like through open router for cron jobs and subtasks. Several members in our community have been asking about running local models, either because the cost or because they want to run their open weight models on their own or they're in countries where API access is a little tricky. So, this Ollama integration is huge for those people. The setup flow now walks you through everything, model selection and the configuration, the entire thing. So, what the interesting thing is is they're positioning this as either a fully local hybrid, so you could run your own main reasoning on something like Ollama 3.1 locally if you have horsepower, but still in the APIs for specific tasks where you would need that extra capability. So, that's actually a pretty smart approach, local

3:03for the bulk work, cloud for the precision stuff. I'm probably just going to stick with my API setup because Opus is just so good and I've tried those other models and just things end up breaking and I don't have a ton of time to waste. So, and I'm not really cost constrained on the agent work I'm doing and, you know, if you have money saved up to throw at it, be my guest. But if you're running hundreds tasks a day and you're just getting started and you want to keep costs down, this local option is going to be a good one. Speaking of other models, there's a new feature that I haven't tried yet, but I'm really excited about this one is the multimodal memory. It's opt-in, so you have to enable it, but it basically allows your agents to now remember images and audio, not just text

3:48conversations. I do a lot of audio transcribing and this is going to be really helpful down the road. So, it uses Gemini embeddings for the indexing. Think about what that means. Your agent could remember a screenshot you showed it 3 weeks ago, or an audio message. Right now, when I'm working with Atlas, the coordination agent that I have in my mission control, or if I'm working with Theriza, who handles my daily content pipeline, they only remember our text conversations. But, imagine if Theriza remembered the screenshots I showed it of the engagement metrics, or like some audio notes I sent through Telegram. I haven't tested this yet, but I literally just updated it an hour ago, and that's probably the one I'm most excited about, but it's on the list of several other features that I've just rolled out. The implementation sounds pretty solid with

4:34this release. They're using Gemini's multimodal embeddings, which are pretty good, and it's opt-in, so you're not paying for indexing you don't need. So, for the Mac users, that's me, and obviously, since I'm running this on a Mac mini, there's a new chat model picker right in the UI. You can switch models on the fly in the chat interface. Here's the cool part, and the thinking level setting persists across all the app relaunches. So, you know, if you've got to set it to show reasoning, it stays that way. This actually is more useful than it sounds. Sometimes I want to see what Atlas is thinking when it's planning out mission control stuff. Other times, I just want a clean output when I'm testing quick interactions. Being able to flip that setting and have it stick is pretty nice. Small quality

5:21of a life thing, but it does add up. They're also completely redesigning the whole iOS home canvas, new welcome screen within the live overview of your agents, docked toolbar. That's pretty nice. I don't use the iOS app much, because I'm usually just at my desk working with the agents directly, but I know some people in Shipping School have been using it in the mobile interface for like quick check-ins, checking on their agents, what they're doing. And the new design does look cleaner from what I can see. Now, here's the thing that directly ties into what I was talking about in the boot camp call we had just a few hours ago. Open router just added two new free stealth models, Hunter Alpha and hunt and healer Alpha. I don't get it with the names, but that's what they're called. They're free

6:07for about a week and then they're probably going to go to normal pricing. Something to consider if you're trying to keep costs down. And this is perfect timing because I was literally telling people today to use Opus as their main model and cheaper models through open router for subtasks and cron jobs. Well, free is pretty much as cheap as you can get. And if you're running daily automation like I do with Ghostface pulling the intel at 5:00 a.m. or Rizza handling the content pipeline stuff at 6:00 a.m., you could potentially route some of those tasks to these free models. I haven't tested the quality yet, but for basic cron tasks or using Brave API to do some web scraping, automating social media posts, or just data processing stuff, they might be pretty good. And free means you can experiment without burning through your API budget. And let me talk about where

6:52I actually hit issues. ACP session resume, that's let's you pick up Codex conversations where you left off. I haven't tried this yet. I just updated, like I said, but it does sound pretty useful for longer coding sessions. You know when you're working on something complex, the conversation gets long, and you want to start fresh, but keep the context? That's what that's for. The thing I definitely ran into was the cron breaking changes. They've made delivery isolation tighter, which is good for security stability, but it means some existing cron setups might break. I had to run open claw doctor to fix after updating, and it did catch a few things. Actually, this relates to what I was just debugging last night. I was having issues with isolated sessions

7:37not spawning properly when I manually triggered them. Wasn't sure if it was my setup or the version I was on. It turns out updating to 311 and running the Claude doctor command fixed pretty much everything. So, if you're having weird cron behavior, the update might actually have solve that issue. And I think we all experience this as we're getting this set up. Cron jobs misfire, maybe you're communicating in the Telegram app, and this just causes a lot of confusion. They've also fixed the launch agent restart issues on Mac OS. This was annoying, and sometimes when I'd restart the agents, or update the configuration, the launch agent wouldn't restart cleanly. You'd have to actually manually go in and kill the process, or reboot. The fix should make that more reliable. There's also token stripping fixes for

8:24GLM 5 and deep seek models, iOS recognition improvements as well, Discord thread auto archive. None of these directly affect my setup, but they do matter for different use cases. Like, if you're communicating with Discord, that's great. Quick aside, if you're trying to build your own agent, and you want to build a team like what I've got running, we call cover all of this inside of our community. We have almost 50 members in just 3 days. We do four live calls a week. We got a starter kit that walks you through everything. We have a code Claude code course, and an open Claude course, as well as just work with you side by side to get the items built. So, we walk you through like model selection, agent architecture, the cron job setup, and pretty much the whole pipeline. I'll put the link down

9:10below. So, what I'm most interested about this latest release after just came out an hour ago, is the multi-modal memory. My content pipeline generates a lot of visual assets. So, screenshots of metrics, exported charts, design mock-ups. Right now, I have to describe those to my agents, or convert them to text somehow. If Risen could just actually remember and reference those images directly, that does change the workflow significantly. Same thing with the audio is like I can record a lot of voice notes now when I'm away, you know, and I can just be on Telegram and just, you know, shipping ideas right to Open Claw. So, ideas for like content or like feedback on campaigns, things like that. For my agents could index and remember those audio files, that's basically giving them access to a much richer

9:56context about my thinking and decision-making process. The Ollama integration is also interesting from a community perspective. I keep getting questions about running local models. So, cost is obviously part of it and also security. Some people want to run hundreds of tasks per day, but the API costs will kill you. And others are worried about sending sensitive business data to external APIs, so local models do solve that problem. What I like about their hybrid approach is you're not locked into one or the other. You could run your daily automation locally, but still hit Opus when you need that extra reasoning capability. So, you kind of get the best of both worlds. The security fix though, that's the real story here. Everyone who gets Open Claw going is all worried about the security.

10:41They're doing more and more with each release update and really listening to the community. Cross-site WebSocket hijacking is serious. If you're running Open Claw anywhere that's not completely isolated, you'll need to do this update. And honestly, even if you're thinking about running locally, double-check your setup. I have a cron job fire off every day to double-check security. Make sure you're not accidentally exposing ports or running behind a local proxy that could leave you vulnerable. So, for my setup, it's not an issue because everything runs on the Mac mini behind me and I can access it locally. So, no reverse proxy, no external access, nothing has been exposed to the internet. But I know people who are running more complex setups, VPS instances, Docker containers, remote access configurations. Those setups need

11:27more security and that helps with this latest patch. The model picker in the macOS chat UI is actually going to be useful for testing. I usually stick with Opus for serious work, but when I'm experimenting with different approaches or trying to optimize costs for specific tasks, being able to swap models quickly can be really helpful. And having the thinking level persist means I don't have to reconfigure it every time I restart. One thing I want to test is how the free open router models perform for specific cron tasks. Ghostface runs every morning at 5:00 a.m. pulling intelligence from various sources. It's mostly data processing and summarization. Riza handles the content pipeline stuff at an hour later at 6:00 a.m. Template generation, formatting,

12:12and basic analysis. These might be perfect candidates for those free models. If Hunter Alpha or Healer Alpha can handle these tasks reliably, then that could significantly reduce my monthly API costs. Even though I'm not super cost constrained, saving money on routine tasks means I can use it more within my budget for more complex reasoning. You know, the stuff where Opus really shines. So, the ACP session resume feature is interesting for longer development sessions. I don't do a ton of coding directly through Openclaw. That's what I use Claude Code for here in the terminal. But when I'm working on Mission Control dashboard of like debugging agent configurations, those conversations can get pretty long. Being able to resume instead of starting fresh every time would be pretty helpful. What

12:58I'm really curious about is how well the multimodal memory works with my existing agent team. So, Atlas, the coordination agent, handles Mission Control and strategic planning. So, if it can remember visual dashboards and charts and screenshots and performance metrics, that would make the planning sessions much richer. Instead of just me describing of what I'm seeing, I could just show it. Same with Inspector Deck, which handles my X posts. Right now, I describe the kind of visual content I want it to reference or include. I end up taking screenshots, and it just ends up becoming a big mess. So, if I could actually have it remember and reference images from previous conversations, that opens up a whole new possibility for content creation and re-engineering the entire pipeline. The current delivery

13:43isolation changes are actually a good thing long-term, even though they caused some breaking changes. That tighter isolation means more reliable execution and better security. My agents pretty much run on complex workflows. The Rizza pipeline alone involves multiple steps, multiple API calls, file generation. And all that's pretty much handled by Opus 4.6. So, the more isolation means less less chance of one task breaking and interfering with another. Running Open Claude Doctor Fix has been helpful. It caught a few configuration issues I didn't even know I had. It's always worth running after any major update as well. The diagnostic tool has gotten pretty good at identifying and fixing common problems automatically. The launch agent fixes on the Mac OS are

14:28going to make it day-to-day operations much smoother. Restarting agents, update configurations, debugging issues, all of which should be a bit more reliable now. Less manual process killing and fewer reboots needed. So, looking at this at the bigger picture, this release feels like Open Claude's maturing as a platform, trying to keep up with Anthropic. First-class local model support, better security, more reliable operations, richer memory capabilities, which has been a common issue for a lot of people. These aren't flashy features, but they are the foundation you need to get some serious agents deployed. For anyone in my community running production agent workflows, this update is worth it just for the stability improvements and security fixes. The new

15:13features are a bonus. If you're just getting started with Open Claude, the improved Ollama onboarding makes local models much more accessible. You can start experimenting without any API costs, figure out what kinds of tasks work well, and then decide if you want to upgrade to cloud models for specific use cases. The free open router models are interesting for testing as well. You could prototype an entire agent workflow without spending anything, then optimize your model selection once you know what works. I'm planning to spend some time this week testing the multimodal memory with my existing agents and just see how well that works with the screenshots and the charts and maybe some audio notes. Maybe it's a benefit, maybe it's not. So, if it works well, I could change how I structure my agent interactions. Also,

15:58want to test some of those free stealth models for some of the cron tasks. Right now, I use Gemini Flash 2.5, costs basically nothing. So, if they can handle the routine stuff reliably, I probably will route some of that work through them while they're free. Good way to evaluate their quality without any costs, I guess. The model picker in the UI chat will be useful for that testing as well. So, quick way to compare how different models handle the same task. Better than switching between different API configurations, I guess. Overall, this is a pretty solid update. The security fix alone makes it worth updating immediately if you're on any kind of exposed setup. The new features are nice additions, especially if you've been wanting to try more of the local models or work a lot with visual and audio content. For my specific setup, the cron fixes and the macOS

16:44improvements are going to make daily operations a bit smoother. The multimodal memory can be a difference maker if it works as advertised. And the free models give me some new options for cost optimization. If you want to build your own agent team and learn this stuff hands-on, then check out the link down below. We have a fast-growing community. We cover all the technical details, help you avoid some of the common mistakes, and get you going and hit the ground running. I do believe being in part of a community with other builders can really help you take the next step forward in your journey from shipping an app or learning some type of automated system. That's what we're building every single day. So, I'll put the link down below. Leave any comments. What are your thoughts on this new update? And if you haven't got subscribed yet, what are you waiting for? Hit the subscribe button.

17:29We'll see you in the next one. Be blessed.