OpenClaw 3.31 Update: Flows Command, Failover Fix, and MCP Remote

By Beau Johnson··8 min read

Quick answer

OpenClaw 3.31 adds a unified background task ledger with the openclaw flows command, treats a generic Anthropic API error as transient so retries and model failover work, blocks dangerous skills by default, supports remote MCP servers, and lets agents search each other's session history.

In this video I go through the OpenClaw 3.31 release and pull out the five or six changes that matter if you actually run agents, so you don't have to dig through a changelog with 30 bullet points. It's for builders, especially people without a software background who depend on OpenClaw to run their business. I recorded this in March 2026. Tools like this change fast, so check the current release notes and docs before you act on anything here.

For context, I run 17 AI agents on OpenClaw every day. They power my content pipeline, and I've named them after the Wu-Tang Clan: the Rizza writes video scripts and outlines, Inspectah Deck handles X posts, Ghostface gathers intel, Method Man cuts clips, and U-God makes thumbnails.

What I cover in this video

  • 1:30 Background task control, and why tracking tasks used to be a mess
  • 2:15 The new openclaw flows commands and the SQLite task ledger
  • 3:01 The Anthropic failover fix, and how an outage took my cron jobs down
  • 6:03 Security hardening for skills, plugins, and gateway auth (breaking changes)
  • 7:36 Remote MCP server support over HTTP and SSE
  • 9:09 Memory search across agents
  • 10:40 Quick hits: WhatsApp emoji reactions, Slack exec approvals, QQ bot support
  • 13:42 Should you upgrade? How I approach it

Background task control with the flows command

This is the biggest change for me. Before 3.31, cron jobs, sub-agents, ACP tasks, and background CLI processes were scattered. If you wanted to know what was running, what failed, or what got stuck, you had to check several places. In 3.31 all of that sits in a single SQLite backend ledger, and there's a new set of commands to work with it:

openclaw flows list
openclaw flows show
openclaw flows cancel

Now one command shows every background task in one place. Lost-run recovery is built in, so a task that crashes or gets interrupted can pick back up. There's also auto cleanup, so dead "zombie" tasks don't sit around using resources. I've had cron jobs fail and didn't find out until I checked the output hours later. If you run more than three or four agents, I think this feature alone is worth upgrading for.

The Anthropic failover fix

I ran into this one the morning I recorded. Anthropic had an API outage, and the API kept returning a generic "an unexpected error occurred" message. In 3.30, OpenClaw didn't treat that as a transient error. So it didn't retry and didn't fail over to another model. It just died. Most of my cron jobs run on Claude, so my whole morning pipeline stopped. The Rizza didn't write outlines, Inspectah didn't post, Ghostface sat idle, and everything was down for a couple of hours.

3.31 treats that error as transient. OpenClaw retries, and if the retries fail, it falls back to your backup model. My failover model is GPT-5.4, so if the same outage happened again, my agents should switch over and keep running. That matters for anyone running production workloads.

Security hardening and the breaking changes

Pay attention to this one, because some of it breaks existing setups:

  • Dangerous skills and plugins fail by default. If the code has dangerous findings, the install is blocked. Before, it installed anyway and you had to trust it.
  • Stricter gateway auth. Trusted proxy configs are tighter, local callers need real tokens (being on the same machine isn't enough anymore), and node commands stay disabled until pairing is approved, not just started.
  • ACP approvals. Tools that can run code indirectly through ACP need explicit approval every time.

I already run VirusTotal scans on skills, and I often just give my agent a skill's link and have it rebuild the skill instead of installing it. I mention that there are about 35,000 skills on ClawHub and roughly 3,000 of them have viruses, so this new layer helps everyone, especially beginners and anyone running agents for clients or a team. The breaking part: if you have plugins or skills that were allowed before despite dangerous code findings, they may stop working after you upgrade. Check your plugins first.

Remote MCP servers and cross-agent memory

MCP (Model Context Protocol) is how your agents connect to outside tools and services. Before 3.31 you could only connect to local MCP servers. Now you can connect to remote ones over HTTP and SSE, with auth headers and credential redaction so your API keys don't show up in logs. Remote MCP servers are popping up for databases, APIs, search tools, and more, and you can now add them with a remote server config. No workarounds needed.

Cross-agent memory search is a quieter change, but a strong one. Each agent used to be able to search only its own past sessions. Now you can configure agents to search each other's session history. If Ghostface gathered intel on a trending topic three days ago, the Rizza can find it while writing a script. Right now most of my agent coordination goes through a database (write to a table, read from a table), which works but is clunky. This gives agents a more natural way to build on each other's work. More on multi-agent setups in my AI agents hub.

Smaller quality-of-life updates

  • WhatsApp emoji reactions: agents can react with an emoji, like a heart on a photo, instead of always typing a reply.
  • Slack exec approvals: approval prompts stay in Slack instead of sending you to the web UI or terminal.
  • QQ bot support: a bundled QQ bot channel plugin with multi-account setup, slash commands, reminders, and media support.

Should you upgrade?

I was still on an older build when I recorded this, and I don't upgrade on every release. Only when the features would really improve my setup. This is the first release in several weeks that clears that bar for me. My business runs on this pipeline, though, so I let a new version settle for a few days, read the feedback, watch the GitHub issues for reported problems, and then upgrade on a quiet day. The failover fix makes me want to move faster, but I'm still taking that approach.

Key takeaways

  • openclaw flows list, show, and cancel give you one view of every background task, plus recovery and auto cleanup.
  • 3.31 treats the generic Anthropic error as transient, so retries and failover to a backup model actually work.
  • Dangerous skills and plugins are blocked by default. Check your installed plugins before upgrading, because some may stop working.
  • Remote MCP servers over HTTP and SSE, plus cross-agent memory search, open up more for multi-agent setups.
  • Don't upgrade out of FOMO. Let a release settle, read the breaking changes, and upgrade when you're ready.

FAQ

What does the openclaw flows command do?

In OpenClaw 3.31, openclaw flows list, show, and cancel let you see and manage every background task in one place, including cron jobs, sub-agents, and ACP tasks. It runs on a single SQLite ledger with lost-run recovery and auto cleanup.

Does OpenClaw 3.31 fix failover during Anthropic outages?

Yes. In 3.30, the generic 'an unexpected error occurred' response from the Anthropic API wasn't treated as transient, so jobs just failed. In 3.31 OpenClaw retries that error and then falls back to your backup model if the retries fail.

What are the breaking changes in OpenClaw 3.31?

Skills and plugins with dangerous code findings are now blocked by default, so ones that were allowed before may stop working. Gateway auth is also stricter: local callers need real tokens, node commands wait for approved pairing, and tools that can run code through ACP need explicit approval.

Should I upgrade to OpenClaw 3.31 right away?

Beau recommends waiting a few days, watching GitHub issues for reported problems, checking your plugins, and upgrading on a quiet day. If your setup already works and you don't need the new features, you don't have to upgrade.

Chapters

  1. 0:00Intro
  2. 0:45Why this update matters
  3. 2:15Background task control plane (flows command)
  4. 5:00Security hardening (breaking changes warning)
  5. 7:30Anthropic failover fix (saved my crons)
  6. 9:45MCP remote server support
  7. 11:30Memory cross-agent search
  8. 13:00WhatsApp emoji reactions + Slack approvals
  9. 14:30Breaking changes you need to know
  10. 16:00Should you upgrade?

Transcript

Read the full transcript

0:00Openclaw 3.31 just dropped and here is everything you need to know. As if the news AI cycle wasn't enough today, we got the Openclaw team shipping and this one's a big one. Here's everything you need to know. I run 17 AI agents on Openclaw every single day and I am a very heavy user on this platform and I'm going to tell you what exactly matters in this latest update versus what is all the noise. We don't want to get caught up in that. We want to get actual useful practical tips from this latest feature release. So, when you look at the change log with like 30 bullet points, your eyes glaze over. You don't know what's important. So, I went through the whole thing and pulled out five or six features that actually matter for

0:45builders like you and me. These are generally going to change how your AI agents operate and if you're new here, my name is Beau. I run the Build in Public channel. We went from 1,000 subscribers to 8,000 subscribers in just 21 days. Thank you for subscribing and that was not me. That was my AI agents powering this whole content engine. So, we got the Wu-Tang Clan running this whole thing. You know, we got the Rizza writing the video scripts and outlines and Inspectah Deck handles all the X posts. Ghostface handles all the intel gathering. Method Man chops everything up into clips. We got Unit U God generating all the thumbnails and we have 177 members joining our school community in the last 18 days. Thank you for being here. We basically just tell teach people how to take their next step in AI. We do now seven live calls every

1:30single week. And these are real people building real projects and we're all using Openclaw to do so. So, you can check that out link down below. So, if you're reading the change log and maybe you're not a developer or maybe you're reading it as someone who depends on this thing to run your whole business and you don't have a software background. That's why I make these videos to help just regular people like you and me understand how we can implement some of these features. The first one and this is the biggest one for me is background task control. So, before 3.31, if you had cron jobs running, then sub agents doing their thing, ACP tasks, background CLI processes, all of that was kind of scattered. You couldn't really see what was going on all in one place. You'd have to check different

2:15things to figure out what was running, what failed, what got stuck. It was a mess. Very overwhelming. Now, in 3.31, they unified all that into a single SQLite backend ledger. So, they gave us this new command, Openclaw flows list, Openclaw flows show, Openclaw flows cancel. So, now you can see every single background task your system is running. Cron jobs, sub agents, ACP, all of it. And it's all in one place in one command. And it has lost run recovery built in. So, if something crashes or gets interrupted, it can pick back up. Auto cleanup, too. So, you're not going to have zombie tasks sitting there eating up resources. For someone like me running 17 agents with cron jobs hitting

3:01every couple hours, this is a massive upgrade. I've had situations where cron jobs would fire off, something goes wrong, and I don't even know about it until I check the output later hours later. So, now I can just run Openclaw flows list command and see everything, which is pretty cool. So, if you're running more than like three or four agents, this feature alone is worth the upgrade. This is the kind of infrastructure you need when you're running agents at scale. It's just plain and simple. The second thing and I need to talk about this one because it literally happened to me this morning, the Anthropic failover fix. So, here's what went down. Earlier today, there was an Anthropic API outage. There actually is still an outage going on right now. It's kind of becoming the norm. Real quick, before we keep going, if you're

3:46watching this and you want to actually build with some of these tools, not just watch videos about them, then you're going to want to check out our community down below, Shipping School. We have a full Claude code course, a full Openclaw course and [music] four live bootcamps every single week where we actually help you get set up from scratch. Like actually set this thing up, not just watch a tutorial and figure it out by yourself. And we also provide one-on-one coaching, so you can book a call with me, we can share screens and I can help you get Claude code or Openclaw running on your machine. That's it, no fluff. I built this community because watching YouTube's only gets you so far. We launched it just 3 days ago and we have over 55 members. You need people around you who are actually building, people

4:32who hold you accountable and coaches who can help you when you get stuck. I'll put the link in the description down below. Get in now before the price goes up. Transient errors, the API was returning this generic and an unexpected error occurred. And in 3.30, the earlier build of Openclaw, it did not treat that as a transient error. So, it didn't retry. It didn't failover to another model. It just died. So, what happened to me, all of my cron jobs that run on Claude, which is most of them, they all failed. So, my morning pipeline did not fire off, the Rizza didn't write the outlines, Inspectah didn't do anything, Ghostface was just sitting there chilling doing nothing, and the whole operation was down for a couple hours until the API came back and then also was down again as I'm making this video.

5:17But, I was sitting there and you know, trying to figure out what to actually put into place and figure out what a good failover method was. And this failover is supposed to handle this with this newest update. The 3.31 fixes that. You know, the specific error message is is now treated as transient, so Openclaw will retry it and if the retries don't work, it will failover to your backup model. I have GPT-5.4 as the failover model. So, in theory, if the same outage happens again tomorrow, my agents would have just switched to GPT-5.4 and kept going. I would have barely noticed. And that is a huge deal for anyone running production workloads. This is the fix alone. This one fix has me seriously thinking about when to

6:03upgrade because I felt the pain today. Like I said, sometimes you don't appreciate the fix until you've personally gone and burned through something that's been fixed in a later release. This is the bug that I had issues with this morning. You know, the third thing is security hardening and this one comes with someone breaking changes, so pay attention. In 3.31, if you try to install plugin or a skill that has dangerous code in it, it now fails by default. Blocked. It won't install it. Before this, it would install anyway and you just had to trust that the code was safe, which you know, that's not great. That's kind of a scary one. They also tightened up the gateway authentication. So, trusted proxy configurations are stricter now. Local callers need real tokens, no more

6:49relying on just being on the same machine, and node commands are disabled until pairing is actually approved, not just initiated. Any tools that can indirectly execute code through ACP now require explicit approval every time. No shortcuts. For me, I already run VirusTotal scans on every skill and I also just give the link of the skill to my AI agent to recreate it, so there is no, you know, backend scarcity for me if I'm installing random skills. Keep in mind, there's about 35,000 skills on ClaudeHub right now and about 3,000 of them have viruses. So, be careful. This update breaks that. You know, it breaks the kind of caution that is in this platform. You know, with this new update, there's a whole new security layer for skills, which is great for the

7:36entire community, especially for beginners because not everyone is as paranoid as me. And if you're running agents as a service or for clients or for a team, this is critical. You don't want some random skill running arbitrary code on your machine. That's how bad things happen. The breaking change part, if you know, have plugins or skills installed that were previously allowed despite having dangerous code findings, these might stop working after you upgrade. Just a heads-up. Check your plugins before you update. Don't just blindly upgrade and then just I just don't want you making that mistake. The fourth thing is MCP remote server support. So, MCP is model context protocol. It's basically how your agents connect to external tools and services.

8:23Before the 3.31 update, you could connect to local MCP servers. Now, you can connect to remote ones over HTTP and SSE. So, they added auth headers and credential redaction, too, so your API keys aren't just floating around in the logs where someone can find. Why does this even matter? Because the MCP ecosystem is growing pretty fast. There are remote MCP servers popping up for all kinds of things, databases, APIs, search tools, basically everything now. And now, Openclaw can connect to all of them native. There's no workarounds, no hacks. Just add the remote server config and go. This opens up a lot of possibilities for what your agents can do. I think we're going to see some really creative use cases come out of

9:09this in the next few weeks. People are going to build some wild stuff with this, which I love. People keep building and shipping and I love the community. I love where it's headed. Things have been changing quite a bit, haven't they? The fifth thing is memory cross agent search. This one is subtle, but super powerful. Before, each agent's memory was terrible. Atlas, my main agent, could only search Atlas's past sessions. Could only search Rizza's past sessions. They were totally isolated from each other. But, with this new update, 3.31, you can configure agents to search across each other's sessions histories. So, if Ghostface gathered some intel 3 days ago about a trending topic, the Rizza can now actually find that and

9:54reference it when they're writing a future script or outline. They can share contacts without me manually copy pasting stuff between them. And that is a huge shift for multi-agent setups. For my operation with 17 agents, this is a huge one. Because right now, a lot of the coordination between agents happen through the database, which works, but it's a little clunky at times. You got to write to a table, read from a table, and this gives them a more natural way to build on each other's work. I'm really excited to experiment with I upgrade. I am still on the older version. I had to come out and make this video, and this one is worth upgrading to. And that's a point I should make to people who maybe have their system working already. Don't upgrade to an open claw, you know, new version if

10:40there's no features in there that you could use. Not every release that comes out, I don't upgrade every single time. Only if it's important enough that it actually drastically changes my setup in a positive way. So, don't feel like you need to update every single time. Only if the features are worth you implementing. You know, just in my opinion, for my use, I think it's going to make my whole pipeline smarter, so therefore prompting an upgrade. The sixth thing, I'm going to do a rapid-fire on a few smaller features that are super important. WhatsApp emoji reactions, your agents can now react to WhatsApp messages with emojis instead of always just sending a generic text reply. So, if someone sends a photo, your agent can just drop a heart react instead of typing, "Nice photo." It's a small thing, but it does make the agent

11:26interactions feel a little more natural and human. So, less robotic. I love it. Also, Slack exec approvals. So, if you're running OpenClaw and your team uses Slack, approval prompts now stay at Slack. Before they'd fall back to the web UI or terminal, which meant you had to leave Slack to approve something. Now, you can just stay right where you are. For teams, for example, this is a quality of life improvement that's going to save you a lot of context switching. And context switching is what kills productivity. We all know that. If you're struggling to keep up with content, why don't I tell you about to save you about 40 days worth of work? I built something called Content Machine. It's [music] 10 AI agents that run on the OpenClaw orchestration, and they handle everything. Scripts, thumbnails, X posts, blogs, outreach, clips,

12:11newsletters, [music] all of it. So, I went from 1,000 subscribers to 4,000 subscribers on YouTube in 7 days using this exact system. Every single morning, I wake [music] up and the content's already done. I spend maybe 15-20 minutes reviewing and approving them, and I move on with my day. It works for any niche, fitness, finance, real estate, marketing, whatever you are building, and it is 100% completely customizable to your use case. So, you get the mission control dashboard, all of the cron jobs, everything I've built over the last 40 days helping me gain more and more people to subscribe and join the community. So, you plug in your own thing, and it molds it to you. It learns how you talk, and it writes, so it

12:56doesn't sound like AI slop. $97 one-time, it's not a subscription. I'll put the link down below, and you'll thank me later. And QQ bot support, if you're in the Chinese market or you have users on QQ, OpenClaw now uses a bundled QQ bot channel plugin for multi-account setup, slash commands, reminders, media support. This one's huge for international adoption of the platform. OpenClaw is going global, you know, and it's good for everyone. I'm going to address the elephant in the room, you know, and going back to if I'm actually going to upgrade, I'm not going to do it right now, but I am going to do it in the future. I'm still on an older build. Here's why I'm really hesitant is my whole system just works, and I feel like

13:42getting it to the point of your OpenClaw working is a huge victory. So, I have been hesitant to update. This is going to be the first update that I do in several weeks, and I feel like the features are there now worth me prompting to upgrade. So, you know, my whole content and my whole business runs on this OpenClaw setup, and that's why I've been hesitant. You know, our school community, for example, we have 177 members, and our YouTube channel just hit 8,000 subscribers. And all this happened in just a little over 3 weeks. I can't afford to have the pipeline go down because of an upgrade going wrong. You know, that's money on the line. But, my approach is always the same, you know, let a new version bake for a few days, get some good feedback, watch the GitHub issues, and see if anyone reports problems, and then upgrade on a quiet

14:28day. You know, you never know, they might ship an update tonight because they didn't know what had broken, you know, in this new build. So, you know, that's how I'm approaching it, and that's how you should approach it as well. If you have your OpenClaw setup working, you know, give it a few days. That Anthropic failover fix is tempting, I will have to say. I do have something similar going, but this is makes it automatic. That's huge. You know, like I said, I literally felt that pain this morning. So, I might move faster on this one. You know, I'm still going to let it go through its course, and go through those kind of healing moments in this latest build, and and then upgrade. So, you know, if you're running a production setup like mine, basically, I just say I recommend the same approach. Don't rush, read the breaking changes, check your plugins, and upgrade when you're ready and when you're comfortable. Don't just

15:14FOMO into every single up. That's the move. That's OpenClaw 3.1, the major features, you know, background task control plane, Anthropic failover fix, security hardening, MC remote servers, cross-agent memory search, and a bunch of quality of life improvements. This is a really solid update, probably the most meaningful one in the last few weeks. The OpenClaw team has been cooking, and I love it. If you're building with AI and you want to be around other people who are also building with AI, not just talking about it, come check out the community. There's a reason 177 people have joined in the last 18 days, and we now do seven boot camp calls a week. So, if you're looking to get in the game and you need some hands-on help, that's what we're for. We're focusing on community first and human-to-human interaction. The whole goal is just to help you take

15:59the next step in your AI journey. So, we've got people building Chrome extensions, food trailer businesses, AI content pipelines, literally all kinds of stuff. I'll put the link down below. So, if you found this helpful, make sure you subscribe because I release about four to six videos a day um covering AI news, OpenClaw use cases, and people building amazing things with AI. Hit the notification bell as well, and comment down below. Are you installing this upgrade or letting this one pass? We'll see in the next one. Have a blessed day. Bye-bye.

Ready to start building with AI?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

Keep reading