OpenClaw 5.4 Update: Google Meet Voice Agents and Reliability Fixes
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
In this video I go through the OpenClaw 2026.5.4 release. The headline is voice agents that handle live calls much better, and there's a long list of reliability fixes behind it. It's for anyone running OpenClaw day to day, or thinking about putting an agent into a real business workflow, who wants the release notes in plain English.
I recorded this in May 2026. These tools change fast, so check the current OpenClaw docs and release notes before you rely on any detail here.
What I cover in this video
- 0:45 Google Meet and Twilio voice agents through a realtime Gemini voice bridge
- 5:21 Windows gateway fix and Docker path cleanup
- 6:52 Codex audio transcription routing
- 8:25 Plugin performance and faster gateway startup
- 9:57 Clearer install hints for external plugins, plus secret handling
- 11:28 A new command for auth profiles, and channel fixes for Telegram, WhatsApp, Slack, and Discord
- 15:15 Sub-agent and cron reliability (no more "green failures")
- 16:46 Command approval safety and memory search fixes
Voice agents that behave like a real conversation
The biggest change is voice. Twilio dial-in joins for Google Meet can now speak through a realtime Gemini voice bridge. Until now, voice agents usually felt awkward. The pacing was odd, there was latency, audio seemed to pile up behind the scenes, and if you interrupted the agent it didn't always react the way a person would.
5.4 goes after that directly. It adds:
- Paced audio streaming, so speech comes out smoothly
- Backpressure-aware buffering, so audio doesn't stack up
- Barge-in handling: the queue is cleared when someone interrupts
- Removal of Twilio feedback during realtime speech
Put simply, the agent should answer faster, sound smoother, and stop talking when you cut in. That matters most on live calls, especially sales calls. If an agent takes a minute in Telegram or while writing a script in the background, nobody minds. On a call, half a second is noticeable, two seconds can lose you the lead, and five seconds ends the conversation. Once this works well, an agent isn't just a chatbot anymore. It can join a meeting, listen, and respond.
Reliability: Windows, Docker, Codex, and plugins
Reliability isn't an exciting headline, but it's what separates a demo from an agent you can actually sell as a service. Here's what changed:
- Windows: the gateway's default loopback listener binds differently now, so the dual-stack localhost issue shouldn't stall HTTP requests anymore. Windows temp paths and media attachment handling were fixed too.
- Docker: Docker Compose now pins the container-side workspace and config paths, so macOS host paths don't leak into the runtime.
- Codex transcription: OpenClaw now advertises Codex audio transcription, and by default it sends active Codex chat models to OpenAI transcription instead of passing chat model IDs to the audio transcriber. Codex shouldn't try to transcribe audio with a chat model.
- Plugin performance: model catalogs and plugin metadata get reused on more hot paths, so there are fewer cold scans. Gateway startup also defers some sidecars until after the ready signal and fast-paths trusted bundled plugin metadata.
The transcription fix matters to me because my content system is a chain of agents. One finds the signal, one writes the script, one makes thumbnails, one cuts clips, one publishes blogs and newsletters, and one reviews performance. If an audio file goes to the wrong provider, the whole chain can slow down, break, or fail silently.
Friendlier setup, safer secrets, better channels
When your config references an official external plugin that isn't installed, OpenClaw now points you to the install command. Before, it could make a valid config look invalid. That helps beginners, because the people who succeed with this are usually the ones who get through setup.
Secrets handling improved as well. Auth-profile key and token references are kept, and the plaintext secret values are scrubbed. There's also a new command for inspecting saved per-agent auth profiles without dumping secrets:
openclaw models auth list
On channels, Telegram supports explicit numeric forum topic targets in the message tool and has cleaner long-text streaming. WhatsApp supports explicit channel and newsletter outbound targets. Slack has richer progress rendering with Block Kit drafts and keeps sub-agent replies in the original thread more reliably. Discord now shows degraded transport and event loop starvation signals in status and deep status. The Control UI got polish too: a collapsible cron sidebar, an agent-first session filter, and duplicate heartbeat messages collapsed into one bubble with a count.
Cron, command safety, and memory
For me, the cron change is the biggest one. Cron diagnostics now show a failed tool policy on an isolated run instead of letting a blocked run look green. A "green failure" is one of the worst bugs an agent system can have. Fake automation tells you the job completed. Real automation tells you what ran, which tool was missing, and why it failed. There's also a post-compaction loop guard: if an agent falls into the same tool-call loop after compaction, OpenClaw can abort it instead of burning cycles.
For command approvals, OpenClaw added a tree-sitter-backed shell command explainer and detection for risky command-carrier patterns. When an agent asks you to approve a command, you should know what the command actually does, not just what the model says it does. Most people I work with aren't developers, so the system needs to explain risk in plain English.
Memory got two fixes. Searches across both memory files and wiki pages now keep results from both, so wiki hits with higher scores don't crowd out memory. And active memory now skips the memory sub-agent cleanly when no memory plugin is loaded, instead of throwing confusing errors.
If you're building toward this kind of setup, my AI agents page covers the bigger picture.
Key takeaways
- Voice is the headline. Google Meet and Twilio calls through Gemini now have paced audio, buffering, and proper interruption handling.
- Most of the release is reliability work: Windows, Docker, Codex transcription routing, and faster plugin loading.
- Cron diagnostics now report blocked runs instead of showing them as green, which matters a lot for scheduled agents.
- Setup and security are friendlier: plugin install hints, scrubbed secrets, and
openclaw models auth list. - It's not a dramatic update. It's a stability upgrade that makes OpenClaw easier to trust for real business use.
FAQ
What is the main feature in OpenClaw 5.4?
Voice agents. Twilio dial-in joins for Google Meet can now speak through a realtime Gemini voice bridge, with paced audio streaming, backpressure-aware buffering, and barge-in handling so the agent stops talking when you interrupt it.
Does OpenClaw 5.4 fix Windows and Docker issues?
Yes. Windows got a gateway loopback fix for the dual-stack localhost issue, plus temp path and media attachment fixes. Docker Compose now pins container-side workspace and config paths so macOS host paths don't leak into the runtime.
How do I check OpenClaw auth profiles without exposing API keys?
OpenClaw 5.4 adds the openclaw models auth list command. It lets you inspect saved per-agent auth profiles without printing secrets to your terminal.
What changed with OpenClaw cron jobs in 5.4?
Cron diagnostics now report a failed tool policy on an isolated run instead of making a blocked run look green. There's also a loop guard that can stop an agent repeating the same tool call after compaction.
Chapters
- 0:00OpenClaw 5.4 is here
- 1:10Google Meet voice agents
- 3:00Reliability and Windows fixes
- 5:00Codex transcription routing
- 6:20Plugin performance and setup hints
- 8:40Channels: Telegram, Slack, Discord, WhatsApp
- 10:30Cron and subagent reliability
- 12:00Command safety and memory fixes
- 13:30Why boring updates matter
Transcript
Read the full transcript
0:00Openclaw 5.4 just dropped and the whole headline feature in this update is voice. Our agents are getting voice, we can communicate with them, they can communicate back, and in this video we're going to be breaking down every single headline feature in this release. You need to pay attention because this ecosystem is starting to become more reliable and less like a terminal tool and more like a functional operating system. Which you see Codex, you see Anthropic shipping the same features, and Openclaw is going crazy right now with all of these updates. So, this is the update where Openclaw really starts making progress and catching up. You know, I've seen people using Codex personal computer, I've seen people using Claude co-work, and Openclaw is trying to keep up. I still use Openclaw
0:45every day in my setup and I really love that voice feature because it's starting to feel more agentic, it's starting to feel like less like a terminal tool and less like you need technical skills to get going. You don't. So, the main voice feature they did which is really cool is the Google Voice Meet feature. Basically, you know, you just got a huge upgrade when you are talking about Google Meet or Google Voice, you can now use this in your entire system. So, basically Twilio dial-in joins can now speak through real-time Gemini voice bridge. So, if that sounds technical, it's not. Here's what it actually means. Like before voice agents, you know, they would feel awkward, they would sound terrible, they would sound like AI, but you know, basically you would call in
1:31and the agent would talk, but there was a latency. And that's the whole point is like voice agents just weren't good enough until now. You know, there was a weird pacing, they would interrupt you, it would take, you know, too long to respond. There were moments where the audio felt like it was stacking up behind the scenes. And if you interrupted it, it did not always behave like a real conversation, which is super annoying. But Openclaw 5.4 goes after that exact pain. It adds paced audio streaming, it adds back pressure aware buffering, it clears the barge in queue when someone interrupts, and it removes the Twilio ML feedback during real-time speech. So, they cleaned it up quite a bit. In normal human language, the agent should respond faster, talk smoother,
2:17and stop talking when you get cut off. That matters, especially when you're dealing with customers. Because voice is where agents either can feel magical or completely broken. If an agent is helping you inside Telegram, a little delay is fine, you know, that's you talking to your agent. And if an agent is writing a script in the background, and nobody cares if it takes a minute, right? But on a live call, latency kills it all, especially when you're doing sales. I mean, come on, you can't have that. A half a second is noticeable, even though it doesn't seem like a lot. Two seconds is crazy, like you'll lose the lead. And five seconds will just kill the whole conversation. Real quick, before we keep going, if you're watching 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 community down below, Shipping School. We have a full Claude
3:03code course, a full Openclaw course, and 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 could book a call with me, we could 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 three days ago and we have over 55 members. You need people around you who are actually building, people who 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. And when you're trying to show this to a client
3:49or even teach someone in a community or put an agent inside of a real workflow, that's the stuff that matters. This is why I think the 5.4 update is important for anyone that's wanting to get involved with voice. It's not just an Openclaw adding voice, they're refining it. It is an Openclaw basically making voice feel closer to real-time production use. And once that works, the whole category changes. Because now your agent is not just a chatbot, it can join a meeting, it can talk, it can listen and respond, it can be really part of the room. That is a completely different mental model. Most people think AI agents as something you message. That's not It's not the case here. You type a prompt basically is what people think, like a ChatGPT, it gives you a response and we celebrate. Maybe it runs a tool,
4:35right? Maybe it makes a file, but the future is not that. The future is agents sitting inside the tools we already use. Telegram, Slack, Discord, WhatsApp, email, Google Meet, voice calls, your whole desktop, right? Your business dashboard is another one. That is what Openclaw is quietly becoming, your personal computer AI agent that automates everything. 5.4 pushes that forward. The second big thing in this release is reliability. How many times have you updated Openclaw and it's broken? They're becoming more and more reliable with these updates. I know reliability does not sound like a, you know, amazing headline feature. Of course it's supposed to be reliable, but reliability is the difference between a demo and an actual business or an AI
5:21agent that you could sell as a service. There are a bunch of fixes in here that are basically Openclaw getting harder to break, which is awesome. Windows users got a gateway fix. The default loopback listener now binds to 127.0.1 on Windows. That means the weird localhost issue from dual-stack behavior should not wedge HTTP requests the same way it used to. So, if you are non-technical, you do not care about libuv libuv, you care that localhost works when your agent needs to work. And that is the point. There is also a Docker and Windows cleanup theme running through this release. Windows temp paths got fixed, Windows media attachment handling got much better. Docker compose now pins
6:07container side workspace and config paths, so host paths from macOS do not leak into the runtime. Again, that's not flashy, but it's important to note. If you have ever tried to help a beginner get an agent up and running across Mac or Windows or Linux Docker and or just random VPS setups, this is the stuff that can save you hours of troubleshooting. And this is why I always keep saying the real product is not the model. The model is only one piece to the puzzle. The product is the harness, you know, the runtime, the setup flow, the plugins system, the channel routing even, and the most important part, the memory layer. The way it recovers when something goes sideways, that is where Openclaw is putting a lot of the work. The third thing is Codex media handling. Openclaw
6:52now advertises Codex audio transcription in runtime and manifest metadata. And it routes active Codex chat models to the OpenAI transcription by default instead of sending chat model IDs to the audio transcription. So, that is a mouthful, I know, but here's the simple version. Basically, Codex should not be trying to transcribe audio with a chat model. That is how you get weird failures. 5.4 makes that routing much smarter. For us, that matters because we use Codex heavily for coding and utility work. We use agents to pull videos. We use agents to transcribe. We use agents to set these long-form videos into clips. We also use agents just to process a ton of media in general. Every time the media routing gets cleaner, the whole content machine
7:39gets more stable. This is part that people just miss entirely. When I say I have agents making content every day, I'm not saying one magic model writes everything. I'm saying there's a whole chain of command. One agent finds the signal, that's usually 5:00 a.m. and that's Ghostface scraping X, Reddit, Twitter, whatever. One agent writes the script, one agent makes the thumbnail, one agent turns the long-form into clips, one agent publishes the blog and newsletters, one agent evaluates performance on a Monday and Thursdays. If any piece of that chain breaks because an audio file went to the wrong provider, the whole machine can slow down or break or even worse, fail silent. So, these small routing fixes add up. The fourth big theme is plugin performance. Openclaw 5.4 improves how
8:25model catalogs and plugin metadata get reused. Instead of constantly falling back to cold plugin metadata scans, it can reuse the current workspace compatible snapshot across more hot paths. That means less repeated scanning, there's less startup drag, less wasted work, and more responsiveness. This is one of those updates that most users will not notice directly. They will just feel like the system is less sluggish. The gateway startup also got more performance work. Some sidecars are deferred until after the ready signal. Hot path channel plugin imports were avoided. Trusted bundled plugin metadata gets fast path. Native plugin loading avoids paying the giddy source transform cost unless it
9:11actually needs fallback loading. That is a lot of technical detail, I know, but the story is simple to break it down for you is Openclaw basically trying to start faster, load less junk in the background, and avoid doing expensive work on the hot path. That matters when your agent is just not a toy anymore and you're selling this as a service. Once you have cron jobs, memory channels, plugins, and media tools, basically sub-agents and your dashboard all running, that startup time really does matter. Hot path performance is the key, and it matters most if you're trying to go for a quick setup. Control plane performance is the key as well, you know, and we just don't talk about it enough. Like a month ago, Openclaw was so much slower than it is today. The fifth thing is plugin migration. 5.4 adds better install hints when your
9:57configuration references an official external plugin that is not installed. So, before upgraded configs can make it look like you had an invalid plugin setting, now Open Claw can point operators toward the actual install command instead of telling them to remove valid config. This is a beginner-friendly improvement, and I care about that more than almost anything because Open Claw is powerful, but the setup is still a pain in the butt. The people who win with this stuff are usually the people who survive the setup. Like, once you lay that foundation, have fun building. But, it does take time. So, every improvement that makes the setup error less confusing matters here. If someone has a valid Telegram config or Discord config or an external plugin retry, the system should not make them feel like they did
10:42something wrong. It should say the plugin needs to be installed, right? And here's the next step. That is how normal people actually go through processes. There is also a secret refix in this update. Open Claw now preserves off-profile key references and token references field when the scrubbing provider targets secrets. Basically, it keeps the safe reference metadata without keeping the plain text secret values. And that's the right direction. Agents need secrets, they need API keys, they need OAuth tokens, they need channel credentials, right? But, the system has to treat secrets like radioactive material. It's useful, but super dangerous. The right architecture is not no secrets. The right architecture is controlled references, scrubbed plain text, safe config, and good re- pair behavior. 5.4 keeps moving in that direction, and I love it. There
11:28is also a new command called Open Claw models off list. That lets users inspect saved per agent off profiles without dumping secrets. So, that is exactly the kind of boring command that actually saves real debugging time. Like, if your agent is falling behind because it is pinned to the wrong off profile, you need to see that. But, you do not want to print secrets into your terminal or paste them into a chat. So, this is a practical operator feature. The sixth theme is channels. Telegram got explicit numeric form topic target and support in the message tool. That matters for people running communities or structured groups. WhatsApp also got a support for explicit channel and newsletter outbound targets. If you're struggling to keep up with content, well, I'm about to save
12:13you about 40 days worth of work. [music] I built something called Content Machine. It's 10 AI agents that run on the Open Claw orchestration, and they handle everything. Scripts, thumbnails, X posts, blogs, outreach, clips, newsletters, 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 up, [music] and the content's already done. I spend maybe 15, 20 minutes reviewing and approving them, and I move on with my day. [music] 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
12:58mission 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, [music] so it doesn'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. Discord got degraded transport and gateway event loop starvation signals in status and deep status. Slack got richer progress rendering with block kit drafts. Telegram text got long streaming and got cleaner. Slack sub-agent replies stay in the originating thread better. There are a bunch of channel-specific fixes in here, and this is important because channels are where agents actually
13:43become useful. Keeps it more managed, keeps your conversations clean. And nobody wants an AI agent trapped in a dashboard, right? They want it where the work happens. Telegram, Slack, or Discord, or WhatsApp, whatever. That is the operating system layer. Channels are the department. Telegram might be the command center, Slack might be the teamwork space, and Discord might be your community. WhatsApp, we can't forget about that. That might even be the customer communication. Google Meet might be coaching calls or sales calls. So, Open Claw is really trying to make one agent operate across all of those. That's the bigger picture here. The seventh thing I like is the control UE work. The cron new job sidebar is now collapsible. The chat session picker got an agent first filled filter. Chat controls stay responsive across the phone, tablet, or desktop. Duplicate
14:29heartbeat messages collapse into a bubble with the count. In the UE, can record browser long animation frames and long tasks in the debug event log. Control UE mitigates safer short-lived tickets for assisting media fetches. There are a bunch of polish improvements here. And I know some people hear UI polish, and they think it's secondary, but I do not. It's gotten so much better. For agent systems, UE polish is trust. The interface is noisy, people think agents are chaotic. If a heartbeat message spams the chat, people think it's broken. If a cron screen is cramped, people start managing their animations. If media links expose long-lived off tokens, that's not ugly, but that's a huge danger, especially if you do screen sharing. So, yes, the UE matters, especially if Open Claw is
15:15going to be used by normal businesses, not just terminal people. The eighth thing is sub-agent and cron reliability. 5.4 preserves group child results when direct completion fallback bypasses the requestor agent announced turn. It also refreshes deferred final delivery payloads when same session completion output changes. And cron diagnostics now serve failed isolator run tool policy. This is instead of making blocked cron runs look green. The last one is huge. A green failure is one of the worst bugs in this whole agent system. You think it's working, but it's not. If a cron did not actually have the tools it needed, I do not want it quietly pretending everything is fine. I want the system to tell me what failed because this is the difference between
16:00automation and fake automation. Fake automation says job completed, you're good to go. But, real automation is saying this ran, this tool was missing, here is why it failed, and here's the next fix. Come on now, we know that. And that is how you build trust with scheduled agents. And scheduled agents are the heart of a whole system. Our content machine runs on all cron jobs. Intel gathering, script writing, thumbnails, blogs, X posts, clips, publishing, and the performance evaluation agent. If cron's lie, the business suffers. So, that's the kind of diagnostic work that matters. There is also a post compaction loop guard for tools. If an agent gets into the same tool call loop after compaction, Open Claw can abort instead of burning cycles forever. That is one of those scars you only build after seeing the failure, and
16:46that's why. And like that, mature systems are full of scar tissue. You know, every weird bug becomes a guardrail. The ninth thing is approval and command safety. Open Claw added tree sitter backend shell command explainer for future approval and command review surfaces. It also added detection for risky environment command carrier patterns. It treats post-6 exec and command carrier for inline eval, shell wrapper, and source detection. Again, it's not a flashy user feature, but it does matter when it comes to security because agents running shell commands are powerful, and powerful is only useful if it's inspectable. If your agent asks you to approve a command, you need to know what the command actually does, not just what the model says it does. What it actually does. That is where command explainers are stricter
17:32carrier detection, that's when it becomes more important. And this is especially important for the audience we are building for. Most Shipping School members in our community are not professional developers. You don't need to be. I mean, come on, if you're learning to code by hand right now, what are you doing? Like, come on. The people in the community are founders, creators, business owners, resellers, service business owners, HVAC consultants, roofers, you know? They shouldn't have to become shell experts to use this agent. You know, the system needs to explain a risk in plain English. 5.4 is laying out more groundwork for actually that type of user. There is also a memory fix I really like. Memory wiki searches now preserve represent representation from both corpora and corpus equals all searches, so memory hits do not get starred by the wiki
18:17results with higher numeric scores. I know that's nerdy, but it's super important. If you ask your agent something and it searches both memory files and wiki pages, you do not want one corpus to drown out the other just because of the scoring systems that are different. Memory is only as useful as the retrieval and if it works. And retrieval is not just whether the data exists, it's whether the right thing comes back at the right time. That is the whole game. Another small important thing is active memory skips the memory sub-agent gracefully when no memory plugin is loaded. So, instead of throwing confusing allow list errors, it just handles the missing backend cleanly. That's the good operator experience, especially for new installers. And I love it, you know? This all matters because Open Claw is
19:02moving so fast. I mean, it's like an update every other day. If the update systems are fragile, every release becomes scary. If the update systems can migrate plugins, repair stale records, preserve off profiles, and explain what is missing, then users can actually stay current. That's the difference between a hobby project and actual infrastructure. Now, if I had to boil Open Claw 5.4 down to one sentence, it would be this. Open Claw 5.4 makes your AI agents better at living in the real world. Voice calls are smoother, channels are more reliable, cron failures are easier to see, and your plugins load faster. So, the control UE is much cleaner as well, and Windows and Docker users get fewer weird edge case failures. This is not the update where I would say everything changed overnight, and it's not insane,
19:47but it's a nice stability and reliability upgrade. I won't want to be too dramatic here, and this is the update where the system gets more adult, gets more reliable, and something that you can actually trust to build a business on. I do not need my agent platform to be flashy every week, really. I just need it to be dependable. Come on. So, if you want to actually build AI agents like this in build out workflows and use cloud code and use codex and use open cloud to run your business or build your business, check out the link down below, join our community. We have nearly 200 members in there, and I launched it about 40 days ago. We have unlimited one-on-one calls that we just added to the community where, you know, if you want to share your screen and not join the nine group calls that we do every week, we can help share screens, get your open class set
20:32up, or Hermes agent, whatever you want to do, and we'll walk you through the process and set up your first AI automation. Those calls are unlimited. You could book a million of them if you'd like, and we're happy to help you. So, hit subscribe. We got videos like this coming out every couple hours. We'll see you in the next one. Have a blessed day.
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 SkoolKeep reading
OpenClaw 5.2 Update: Cron Jobs, Context Safety, and Security Fixes
The OpenClaw 5.2 update is about reliability, not flashy features. Cron jobs run in their own lane, Codex and GPT-5 runs compact before they hit the context window, and approvals and secret checks are tighter. It also adds Discord voice guardrails and better model routing.
OpenClaw 5.26 Update: Faster Gateway, Reaction Approvals, Security
OpenClaw 5.26 is a reliability release. It brings faster gateway replies, transcript-backed meeting summaries, steadier Telegram/iMessage/WhatsApp/Discord channels, thumbs-up reaction approvals, better realtime voice controls, and tighter security around untrusted browser and file content.
OpenClaw 4.26 Update: Migrate Claude Code Faster, Add Browser Voice, and Fix Agent Friction
OpenClaw 4.26 is one of those releases that does not just add one shiny feature. It removes friction across migration, voice, cost, privacy, and long running sessions.
