OpenClaw 4.25 Update: Phone Push Alerts, Browser Fixes, and Voice

By Beau Johnson··8 min read

Quick answer

OpenClaw 4.25 turns OpenClaw into an installable PWA, so your agents can send web push notifications to your iPhone or Android when they finish a job. It also fixes Playwright browser automation bugs, adds seven bundled TTS providers plus per-agent wake words and voices, and ships six security hardening fixes.

In this video I go through the OpenClaw 4.25 release and what it changes for anyone running AI agents day to day. The main feature is phone push notifications, so your agents can ping you the moment they finish. I also cover the browser automation fixes, the voice and text-to-speech additions, and the security work. It's for people already running OpenClaw and anyone trying to figure out what agents are actually good for.

I recorded this in April 2026. Tools like this change fast, so check the current OpenClaw docs and release notes before you set any of this up.

What I cover in this video

  • 0:00: The headline feature. OpenClaw becomes a PWA and agents can send push notifications to your phone.
  • 0:47: Why this matters for my own setup of Wu-Tang-named agents.
  • 2:18: How push notifications fit with Mission Control and my approve-and-publish workflow.
  • 3:49: The browser automation overhaul for Playwright on single-page apps.
  • 6:51: Seven new bundled text-to-speech providers.
  • 7:37: Wake words and separate voices for each agent.
  • 9:09: The security and hardening fixes.
  • 12:11: Secondary changes: OpenTelemetry, DeepSeek routing, image flags, plugins, and removals.

Push notifications: agents that run while you're away

4.25 makes OpenClaw a PWA (progressive web app). You install it on your phone's home screen, tap the icon, and it opens like a normal app. The gateway can then send web push notifications to that device, the same way Instagram or X pings you. Telegram could already message you, but this is real push to your iPhone or Android.

I run several agents every day, all named after Wu-Tang members. Ghostface gathers intel at 5:00 a.m. from X, Reddit, Hacker News, LinkedIn, and Facebook groups. The Rizza writes outlines and scripts at 6:00 a.m. U-God makes thumbnails. Inspectah Deck handles podcast outreach and posting on X. Method Man cuts long videos into clips and posts them through Post Bridge. Before this update, I only knew a job had finished if I was watching the terminal, checking the dashboard, or opening Telegram.

Now my phone buzzes when a job is done, whether I'm at the gym, the grocery store, or the airport. I review or approve the work and the agent goes back to it. With Mission Control, my day looks like this:

  1. Agents work overnight.
  2. I wake up to a row of push notifications.
  3. I open Mission Control on my phone and review the videos, scripts, posts, and clips.
  4. I approve the ones I want live.
  5. A publish cron job picks them up and Masta Killa posts them to YouTube and X.

That's the difference between agents you babysit at your desk and AI agents that run in the background of your life.

Browser automation fixes for real dashboards

This part is for builders who run Playwright against single-page apps like HeyGen, ChatGPT, or Claude. I've been building an agent called Cappadonna that turns scripts into AI avatar videos with HeyGen Avatar 5 through HeyGen Studio. I spent about three hours debugging that workflow on 4.24, and several of the bugs I hit are fixed in 4.25:

  • Stable tab IDs after form submits and navigation.
  • Persistent Playwright session reconnection when the CDP connection drops.
  • Iframe-aware role snapshots with cursor-clickable detection, so the agent can see and click elements inside frames without switching context by hand.
  • A new openclaw browser doctor command with a deep option that probes the live snapshot for you, so you don't have to launch Playwright codegen just to learn what the buttons are called.
  • Per-profile Chrome MCP command and args with CDP URL mapping, the multi-profile fix for anyone running several Chrome instances.
  • A retry on the raw CDP websocket handshake before the first command, so the agent doesn't quietly die when Chrome is a little slow to connect.

Voice: seven TTS providers, wake words, and per-agent voices

This release bundles seven new text-to-speech providers, including Azure Speech (full SSML support, Opus and PCM output), Xiaomi (MP3, WAV, and streaming synthesis), ElevenLabs v3, Volcengine/BytePlus speech, a local CLI speech provider with voice-note Opus conversion, and Google Gemini with a PCM-to-Opus path so it works for voice notes too.

The voice features I'll use most:

  • Wake word routing. Each agent can have its own wake word. Say "Rizza" and the Rizza starts. Say "Ghostface" and Ghostface starts. It sounds like a gimmick, but switching agents in one voice session used to be a pain.
  • Per-agent TTS voice overrides in the agents config, so each agent gets its own voice and they don't all sound like the same robot.
  • /tts chat with on, off, or default, so you can turn auto TTS on for one session without changing global settings.
  • BlueBubbles, the iMessage bridge plugin, now sends TTS as a voice memo, so an agent can text you a voice memo in iMessage.

Security hardening and smaller changes

4.25 includes six identity and hardening fixes. Device tokens can't be used outside the scope they were issued for. Redaction patterns now apply to saved transcripts before data hits the disk, and to console and file log outputs too. Tailscale-authenticated operator sessions skip the pairing round trip. Mismatched gateway and plugin versions get caught before something silently breaks. The gateway control UI and websocket now support TLS. Launch agent token rotation ends a stale-token loop I've hit myself on macOS. When an agent holds your keys and accounts, you want a project that ships fixes like these every week.

Smaller changes worth knowing about:

  • OpenTelemetry now has token usage and model call duration histograms labeled by agent, so in Grafana you can see spend per agent. There's also a bundled Prometheus diagnostics plugin.
  • DeepSeek V4 Pro and V4 Flash get new routing options through Venice and OpenCode.
  • Image edits get --background, aspect ratio, and resolution flags in the CLI.
  • Plugin install metadata moves to a persistent index, there's a new openclaw plugins registry inspect command, and agent finalize hooks are supported.
  • Heartbeats, cron jobs, and exec runs no longer get injected into the visible prompt, so sessions are cleaner.
  • Skill Workshop and active memory use your configured default model instead of a hardcoded OpenAI model.
  • Telegram renders native quote excerpts for threaded replies.
  • Removals: the plugins.installs config is gone (direct edits get rejected), a persist plugin option is deprecated with a three-month removal target, and the /tmp fallback is removed.

Key takeaways

  • If you run OpenClaw, my advice is to update, install the PWA on your phone, and spend about 10 minutes setting up web push.
  • Push notifications plus Mission Control mean you can review and approve agent work from your phone, away from your desk.
  • If you automate dashboards with Playwright, the tab ID, reconnection, and iframe fixes remove some of the most common failure points.
  • Wake words and per-agent voices make a multi-agent voice setup much easier to use.
  • The security fixes and 100+ contributions two weeks in a row show the project has real momentum.

FAQ

Can OpenClaw send push notifications to my phone?

Yes. As of 4.25, OpenClaw installs as a PWA on your phone's home screen, and the gateway can send web push notifications to your iPhone or Android when an agent finishes a job.

What browser automation fixes are in OpenClaw 4.25?

It adds stable tab IDs after form submits and navigation, Playwright session reconnection when CDP drops, iframe-aware role snapshots, a new browser doctor command, per-profile Chrome MCP settings, and a CDP websocket handshake retry.

Can each OpenClaw agent have its own voice and wake word?

Yes. 4.25 adds wake word routing, so saying an agent's name wakes that agent, plus per-agent TTS voice overrides so each agent can sound different.

Which text-to-speech providers did OpenClaw 4.25 add?

It bundles seven new TTS providers, including Azure Speech, Xiaomi, ElevenLabs v3, Volcengine/BytePlus, a local CLI speech provider, and Google Gemini.

Transcript

Read the full transcript

0:00Open Claw 4.25 literally just dropped a few hours ago and in this latest update they shipped some major voice features that you're going to want to hear. So, they just dropped this today and the headline thing in this entire release is something that I personally wanted since the first time I started running agents on the entire Open Claw platform. Your agents can now text your phone when they're done. So, not in a metaphorical way. Obviously, if you have Telegram, they've been able to do this, but actually push notifications to your iPhone or Android. And the same way, you know, Instagram or Twitter pushes you the second an agent finishes a job. And then the way they did it is the way they turned Open Claw into a PWA that allows you to install on your phone and on your home screen. So, you just tap the icon

0:47and it opens up like an actual app. And then the gateway can push to your device through a web push. So, that is the fix. If you want to read it in the actual commit, I'll put it linked down below. Check it out. But, why is this a big deal for me? And why should it be a big deal for you? You know, I run multiple agents on a daily basis. Ghostface, which yes, all my agents are named after the Wu-Tang, he's the intel gatherer. He wakes up at 5:00 a.m. and starts gathering intel from X through the API, Reddit, Hacker News, LinkedIn, Facebook groups. Then we have the Rizza, writes the outlines and the scripts at 6:00 a.m. Then we have U-God, which is generating the thumbnails. Inspect Deck does all the things on X like podcast outreach and posting to X. Method Man cuts up these long-form videos into

1:33clips and auto post them through Post Bridge. And before today, the only way I would know any of these finished was either I was sitting at the desk staring at the terminal or I checked the dashboard manually or I opened up my Telegram. Now, the second any of them finishes a job, my pocket buzzes. Good news, right? And I'm at the gym or at the grocery store and I'm at the airport flying to LA. I don't know. It doesn't matter. The agent texts me. I open my phone and I see what it got finished. I'll review it or I approve it. The agent goes back to work and then that's the unlock. That's the difference here between agents being a thing you just babysit at your desk and agents actually being a thing that runs in the background of your life, which is what we're all trying to accomplish, right? So, if you pair this with Mission

2:18Control, which I've been running for 2 months now, and basically now my whole workflow looks like this. Agents will do work overnight. I wake up and then my phone has a row of pushes. I tap into Mission Control on my phone and I review the videos, I review the scripts, the posts, the clips, and everything that got produced, and then I approve the ones that I want to go live. The publish cron job picks them up and then Master Killa puts them on YouTube and on X without me touching the keyboard. That is the actual end state for any of you who have been trying to figure out what AI agents are even good for. They are good for running your business while you're not at your desk. That's the whole point. And in 4.25, this release is the one that finally makes that work on your phone. Real quick, before we

3:04keep 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 out our community down below, Shipping School. We have a full Claude code course, a full open Claude 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 open Claude 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

3:49you 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. So, moving on to the second biggest feature in this release, this is for the builder audience. And this one is going to hit hard for anyone who has tried to run Playwright on top of HeyGen or ChatGPT, Claude, or any single-page application. Browser automation got a real overhaul. So, you know, I have been building Capadona. No one knows about this yet. Which is my Wu-Tang member that takes the scripts and turns them into an AI avatar video using HeyGen Avatar 5. This is actually me, by the way. You know, I'm not an avatar. This is me. This is my real hair. I am not an avatar. But, I'm just saying, I'm just

4:34testing it out. And it runs through HeyGen Studio, right? And the truth is, the last 24 hours I've been stress testing this on every single bug that lives in the gap between Playwright and a real production dashboard. So, tab IDs change after a form submit, the Playwright session drops, the iframe the editor blocks our role snapshot, the dashboard URL pattern changes, and on and on. I could go on. It'll bore you. But, nonetheless. So, when I read the 4.25 release notes this morning, and I saw that they specifically fixed several of those exact bugs, I laughed because I had been going through about 3 hours of debugging with this whole workflow. I'm on 4.24, by the way. But, what landed was stable tab IDs after form submissions and navigation. Why? Fixed

5:19number 46137. Check it out in the change log link down below. But, persistent Playwright session reconnection when your CDP drops. Fixed number 6728. Check it out. Iframe-aware role snapshots with cursor clickable detection, which means the agent can see and click into elements inside of frames without doing manual context switch. A new open claw browser doctor {dash} deep command that probes the live snapshot for you. So, you do not have to launch Playwright code gen yourself just to find out that the buttons are called per profile Chrome MCP command and args with CDP URL mapping are fixed in this latest update, which is the multi-profile fix that

6:04anyone running multiple Chrome instances has been begging for. In the raw CDP websocket handshake retry before the first command, so the agent does not silently die because of Chrome. This was a fraction of a second slow on the first connect. Annoying. So, if you've been doing any browser automation on top of any modern dashboard, this release is the release you need. The exact pain points that I personally hit while building my latest agent Kappa Donna. In the last 24 hours, these are all fixed and the team shipped them within a week. That is what real OSS momentum looks like. The third feature, and this is the one that's more subtle, but really powerful for content creators. This is voice and text-to-speech got a massive

6:51expansion. Seven new bundled TTS providers in one release. What? Azure speech with full SSML support and Opus PCM output. So, Xiaomi Mimel with MP3 and Wave, and then with streaming synthesis. 11 Labs V3 added to the bundled catalog as well. And then Volce engine to byte plus speed speech and a local CLI speech provider with voice note Opus conversation. That is amazing. And then also Google Gemini made an appearance here. Uh PCM to Opus transcoding path, so it works for voice notes, too. In terminal listing, plus in this really cool one, voice wake trigger-based routing. So, you can set up a wake word. So, you can

7:37set up a wake word for each of your agents. Say, you know, Rizza on the mic, right? Rizza wakes up. You say Ghostface, Ghostface wakes up. You just say any of your agents' names and they start working on the thumbnail or whatever they're assigned to do. So, there's a different wake word for every Wu-Tang member, and that's what I'm going to be using going forward. And that is the kind of thing that sounds gimmicky on paper, but if you have ever tried to switch between agents in a single voice session, you know how much friction there is. And now you just say the name, and then per agent TTS text-to-speech voice overrides also landed via agents. That is a bracket bracket {dot} TTS feature, so every agent finally gets its own voice. Rizza

8:23can sound different from U-God, U-God can sound different from Method Man. The personalities actually feel like personalities and not like a chorus of identical robots. Text-to-speech has gotten so good nowadays, it's hard to even tell. There's also a quality-of-life fix here called {slash} TTS chat. You can have it on or off or default. It lets you turn it on auto if you'd like for specific sessions without messing without messing with your global settings. And Blue Bubbles, which is the iMessage bridge plugin, now delivers TTS as a voice memo. So, your agent can literally text you a voice memo in iMessage. I think that's pretty cool. Now, the fourth thing I want to cover, and this is the one that tells you whether the project is being run by serious people or not. 4.25 ships a quiet but heavy security release above

9:09the weekly average. So, what landed exactly? Device token scope containment enforcement. A token that was issued for one scope cannot be used outside of the other scope, so there is a full stop. So, if you give a token narrow access, narrow access is what it has, and there is no escape hatch where the token quietly expands to other endpoints. Configured redaction patterns are now applied to persisted transcripts. So, if you have set up a redaction for things like API keys or the redaction now happens before the data hits the disk, not just before the data hits your screen. Same goes for the console and file log sync exits. Redaction is applied at the actual output point, not just somewhere upstream that another

9:54component might bypass. Tailscale authenticated operator sessions now skip the pairing round trip. So, if you are already on a trusted Tailscale network, you are not asked to repair it every time. Mixed version gateway service prevention, which means you cannot accidentally end up with a 4.20 gateway talking to a 4.25 plugin and vice versa, and the system catches it before something silently breaks. TLS enabled gateway control UI and web socket finally landed in these latest updates, so the control plane gets the same transport security as the agent runs, and launch agent token rotation and stale token mismatch loop ends a class of zombie service failures that I have personally hit on this Mac OS in the

10:40last few weeks. So, if you're running Open Claw and you care about giving an agent access to your stuff. If you're struggling to keep up with content, what I'm about to save you about 40 days worth of work. 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 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,

11:25marketing, 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 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. You're giving the agent a lot, you know, your keys, your gateways, your bots, your accounts, right? >> You want the project that holds these credentials to be run by people who are reading the security issues and actually shipping fixes every single week. 4.25 is that update. Six identity and

12:11hardening fixes in one release. That is a serious release. There's a few secondary things that you should not gloss over, but I do want to mention them. The open telemetry stack got a real expansion. Full gen I semantic conventions, including a token usage histogram and a model called duration histogram. Both labeled by the openclaw.agent dimension. So, if you run a Grafana, right? You can finally see your token spend split by agent. Rizza spend versus Ugod spend versus Method Man spend, all labeled just so I know what agents are using. Plus, signal specific endpoint overrides for traces and metrics and logs. W3C trace parent header propagation and a bundle diagnostic Prometheus plugin with a

12:57gateway scape route. So, if you've ever wanted to run Prometheus against your agent setup, that is now a one-line configuration. Deep Seek version 4 Pro and 4 Flash got two new routing options through Venice and open code. And more options for the same model with reasoning content placeholders preserved. Image generation got a {dash} {dash} background, aspect ratio, and resolution flags exposed to the CLI for image edits. So, you do not have to drop down to the provider native config to control output dimensions anymore. Plugin registry got a overhaul where the install metadata moved to a cold persistent index. There's a new OpenClaw plugins registry inspect command and the four agent finalize hooks are now

13:42supported, which is the hook surface a lot of plugin authors have been asking for. Heartbeat, cron jobs, and exec runs are now treated as transient runtime contexts instead of getting injected into the visible prompt, which means cleaner sessions. Skill workshop and active memory now run on the configured agent default model, so no more hardcoded OpenAI for those flows. And Telegram finally renders native {quote} excerpts for threaded replies. On the remove side, the OpenClaw disable persist plugin got marked deprecated, so 3-month removal target. The plugins.installs config is absolutely gone and migrated to persistent index. Direct edits to that section get rejected. So, the legacy manifest alias

14:27scans got removed from the startup and provider paths. And the /tmp and OpenClaw fallback got removed absolutely. So, cleaner system, smaller attack surface. So, the big picture, why does this even matter, right? That's a lot of fixes. What do we even do with this information? Well, 4.25 literally lands 1 day after 4.25 ship, which was an amazing Google Meet feature integration, and stacks neatly on top of it. So, imagine a Google Meet call where your agent joins it. It listens for the wake word. It replies through one of the seven new TTS providers, then pushes the action items to your phone the second the call ends. That is a workflow that did not exist last week, but it exists now. The phone as control surface unlock with PWA and web push is the piece that

15:14makes this thing that you should check in on the go instead of something that you babysit at your desk. That browser automation reliably fixes our direct response to the explosion of dashboard driven workflows. Those That's what people are wiring up, right? Including the HeyGen avatar pipeline that I'm building right now. And 100 plus contributions 2 weeks running that tells you that the project has a real community momentum going. Not just one person shipping in a garage. My take on all this is if you're running Open Claw, update right now. Install the PWA on your phone. Spend 10 minutes setting up the web push, then walk away from your desk and watch your agents task you when they're done. So, that alone is worth the up. If you're doing any kind of browser automation on dashboards, the Playwright reliability fix is going to make the next thing you build half as

16:00painful as the last. And if you have not been on the fence about joining AI and sitting on the sidelines, you need to get going. You need to start building. You need to start using these tools because you can only watch so many YouTube videos. It's time to get your hands dirty. So, check out the release notes. Like I said, I put it down below in the comments. You know, if you want to see more Open Claw builds and you want to be around other people who are actually shipping instead of just talking about them, we run a community called Shipping School and we do nine live calls a week. We have over 200 members in the community and I launched it over 30 days ago. I'll put the link to it down below and if you haven't subscribed to the channel already, please do cuz I release videos like this every single day. We'll see you in the next one. Have a good

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