OpenClaw 4.22 Update: Streaming Voice Agents, Grok, and /models add
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
In this video I go through the OpenClaw 4.22 release. The biggest change is voice agents, which now work with real-time streaming on several providers instead of only one or two. This post is for anyone running OpenClaw, or thinking about it, who wants to know which changes affect their workflows.
I recorded this in April 2026. Tools like this change fast, so check the current docs before you rely on any detail here.
What I cover in this video
- 0:00: Why voice agents used to feel slow or tied to a single provider
- 1:30: xAI Grok gets image generation, six TTS voices, and speech-to-text
- 3:46: Deepgram, ElevenLabs, and Mistral join the Voice Call streaming pipeline
- 4:31: TUI local embedded mode (no gateway needed) and auto-installing plugins during onboarding
- 6:03: Adding models from chat with
/models add - 6:49: Claude Opus 4.7 on Amazon Bedrock Mantle, plus other provider updates
- 8:19: WhatsApp reply quoting, sessions inbox filters, and Control UI polish
- 12:11: Faster plugin loading and Tokenjuice bundled as an opt-in plugin
Voice agents now work across providers
Before OpenClaw 4.22, a voice agent that listens and replies in real time had two choices: OpenAI or xAI. The other providers only handled batch audio, so you waited for the full file before getting anything back. That meant either an awkward two-second pause after you stopped talking, or being locked into one provider, which gets expensive.
xAI got a big set of changes in this release:
- Grok image generation in a regular model and a pro model. Both can edit a reference image, so you can upload a picture and ask for changes.
- Six text-to-speech voices that sound human. Output can be MP3, WAV, PCM, or G.711, the format phone systems use, which helps if your agent needs to call a real phone line.
- Grok STT for audio transcription, plus xAI real-time transcription for Voice Call streaming. If you have Voice Call set up, the whole voice loop can run on Grok, streaming both ways, and it costs much less than doing the same thing through OpenAI.
Deepgram, ElevenLabs, and Mistral were also added to the Voice Call streaming pipeline, so five providers now support real-time streaming transcription: OpenAI, xAI, Deepgram, ElevenLabs, and Mistral. I think Deepgram is the gold standard for accuracy, especially with accents or a lot of background noise. ElevenLabs also got Scribe v2 for batch transcription on inbound media, so an audio file dropped into a channel can be processed.
I plan to use this for outbound cold calling for our app, The Magic Hand, which targets real estate agents and brokers. If you want more background on agents like this, see my AI agents page.
Easier setup: TUI embedded mode, onboarding, and /models add
TUI local embedded mode. Until now, using the OpenClaw terminal chat meant running the gateway in the background. Now you can run the terminal chat embedded locally without the gateway daemon. Approval gates still work, so OpenClaw still asks before anything executes. This is useful on a laptop-only setup or if you want to try OpenClaw without doing the full setup.
Onboarding installs missing plugins. Previously, if you chose something like OpenAI and Telegram during onboarding, it assumed those plugins were already installed. If they weren't, you got confusing errors later. Now onboarding installs any missing provider and channel plugins during setup, so the first run goes more smoothly, including when you're setting it up for someone else.
Adding models from chat. Adding a model used to mean editing the config file and restarting the gateway. Now you can do it from inside the chat:
/models add
You give it the provider name and model ID. The model is registered and usable right away, with no restart. If you're new to OpenClaw, my OpenClaw page covers the basics.
Providers, WhatsApp, and UI updates
- Claude Opus 4.7 on Amazon Bedrock Mantle can now be called through a proper bearer auth route. Before, AWS bearer tokens were treated like Anthropic API keys. This matters for teams that use Bedrock for compliance or have AWS credits to spend.
- Tencent Cloud is now bundled, and you can onboard through TokenHub. It includes preview models with tiered pricing.
- OpenAI Responses now uses OpenAI's native web search tool automatically when web search is on and you haven't pinned a search provider.
- GPT-5 prompt overlay moved into a shared provider runtime. Calling GPT-5 through OpenRouter, OpenCode, Codex, and similar routes now behaves the same as calling it directly.
- WhatsApp gets native reply quoting, so your agent can quote the message it's answering. Per-chat system prompts for groups and direct messages now apply on every turn instead of only some of the time.
- Sessions and Control UI: the sessions list has inbox-style filters (by label, by agent, and by search), and you can set a browser-local operator name and avatar. Quick settings, agent fallback chips, and narrow-screen chat layouts were also cleaned up.
- Gateway diagnostics: payload-free stability recording is on by default. You can export support-ready bundles with sanitized logs, status, health, config, and stability snapshots.
Speed and stability fixes
- The default thinking level for reasoning-capable models went from off/low to medium if you haven't set it yourself.
- Plugin install now adds new plugin IDs to the plugins allow list before enabling them, which avoids permission errors.
- Memory search moved to sqlite-vec KNN for vector recall, which makes semantic search faster and more accurate.
- A WhatsApp bug that sent the same cron message 7 to 12 times is fixed. It happened when a reconnect fired mid-delivery after the 30-minute inbound silence watchdog.
- On Linux, if the system runs out of memory, short-lived worker processes get killed first and the long-running gateway stays up.
- Status has a new runner field that shows whether a session runs on embedded Pi, a CLI backend, or an ACP harness agent like Codex or Gemini.
- Plugin load time is 82 to 90% faster for bundled plugin modules, and the non-interactive doctor run takes about 74% less time because its plugin passes now load lazily.
- Tokenjuice is bundled as an opt-in plugin. It compacts noisy exec and bash tool output in Pi embedded runs, which lowers token use and keeps the context window cleaner.
Key takeaways
- Real-time voice agents now work with five streaming providers (OpenAI, xAI, Deepgram, ElevenLabs, Mistral), so you aren't locked into one.
- xAI can now cover the whole voice loop: image generation, six TTS voices, speech-to-text, and real-time transcription.
- Setup is easier: terminal chat runs without the gateway, onboarding installs missing plugins, and
/models addworks without a restart. - Enterprise teams can call Claude Opus 4.7 through Bedrock Mantle with proper bearer auth.
- Faster plugin loading, the WhatsApp duplicate-send fix, and the optional Tokenjuice plugin make day-to-day use quicker and cheaper.
FAQ
Which providers support real-time voice streaming in OpenClaw 4.22?
Five providers support real-time streaming transcription in the Voice Call pipeline: OpenAI, xAI, Deepgram, ElevenLabs, and Mistral. Before 4.22, only OpenAI and xAI did.
How do I add a new model in OpenClaw without restarting the gateway?
In 4.22, type /models add in the chat and give it the provider name and model ID. The model is registered and usable right away, with no config edits or gateway restart.
Can I use OpenClaw without running the gateway?
Yes, for terminal chat. 4.22 added a TUI local embedded mode that runs without the gateway daemon. Approval gates still work, so OpenClaw still asks before anything executes.
What is Tokenjuice in OpenClaw?
Tokenjuice is an opt-in plugin bundled with OpenClaw as of 4.22. It compacts noisy exec and bash tool output in Pi embedded runs, which lowers token use and keeps the context window cleaner.
Chapters
Transcript
Read the full transcript
0:00Opencloth 4.22 just dropped and this one is feature packed and in this video I'm going to be breaking it all down because voice agents just got a major upgrade. You know, this just dropped a few hours ago and the big one in here is the voice agents. I want to go over how you can use this and how I'm going to implement it. You know, this actually changes a lot of our workflows because up until now, you know, if you wanted to build a voice agent that can essentially just listen to you in real time and respond without sounding like AI slop and without sounding sluggish, you had two options. And you can use Open AI or XAI and that was it. Everybody else was either on batch only, meaning you had to
0:45wait for the whole audio file to finish before you got anything back and they did not support streaming at all. So basically your voice agent would either feel very slow, like there is this awkward two-second gap between you stop talking basically and when it responds and that's a huge problem in just AI voice in general. There's a lot of latency. Or the other option is you were locked into one provider, which gets pretty expensive pretty quick. So with 4.22, you know, that all completely has changed. I'm going to walk you through everything that shipped in this update. That first one and this is the headline is XAI just got a massive upgrade. We're talking about image generation with Grok
1:30image. There is the regular model and then there is the pro model. Both of them support reference image edits. So you can feed in a picture and say, you know, make this change but cooler and it actually works. On top of that, XAI now has six live voices for text-to-speech. Six of them. And these are not robotic text-to-speech voices. They're actually really good. These are actual human-sounding voices. They output an MP3, wave, or PCM, and even G.711 for phone systems. So, that's pretty cool if you're building something that has to dial into a real phone line. On top of it, they added Grok STT, which is their speech-to-text model for audio
2:15transcription. And they added xAI real-time transcription for voice call streaming. So, now if you have voice call set up in Open Claw, you can now do the full voice loop end-to-end on Grok, and it just works in real time, streaming both directions, and the cost is way lower than running the same thing through OpenAI. 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 out our community down below, Shipping School. We have a full Claude code course, a full Open Claw course, and four live bootcamps every single week, where we actually help you get set up from scratch. Like, actually set this
3:00thing 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 Claw 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 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. But, here's where it gets even better is because they did not just add xAI, they added Deep gram, Eleven Labs, and Mistral to the voice call streaming pipeline as well. So, now five different
3:46providers can do real-time streaming transcription: OpenAI, xAI, Deep gram, Eleven Labs, and Mistral. And Deepgram is the gold standard for accuracy. You know, if you've ever tried to transcribe a call where maybe someone has an accent or there's a ton of background noise, Deepgram is the way. 11 Labs also got Scribe version two for batch transcription on inbound media. So, let's say you drop an audio file into the channel, 11 Labs can now process it and the net result is voice agents are no longer an open AI type thing. And they are no longer sluggish. So, if you've been wanting to build a voice agent, but you were worried about getting locked in, this is the update for you. I can think of multiple ways on
4:31how we can implement this into our workflows. Like for me, I'm going to use it for outbound call cold calling for our app called The Magic Hand. We're trying to target real estate agents and real estate brokers, and this is going to be the direction that I'm going to be heading. So, the second big thing is the two-e. There is a new local embed mode where you can run a terminal chat directly without a gateway. Up until now, if you wanted to use Open Claw terminal, you had to have the gateway running in the background. Which for most people, that's fine, right? But if you are on a laptop or you just want to bang out a quick chat without spinning up the whole service, it was pretty annoying. Now, you can just run it locally embedded and then plug in the approval gates. They all work. So, security is not compromised. You still
5:17get asked before anything executes, but you do not need the gateway daemon. So, it's huge for laptop-only setups and huge for people who just want to try Open Claw without committing to the entire full setup. So, shout out to Fuller Stack Dev on that one. That was amazing. The third thing is, and this one is going to save a ton of people time, is onboarding now auto installs the missing provider and channel plugins during setup. So, before this was a nightmare, you know, if you ran through the onboarding and you said, "I want to use OpenAI and I want to use Telegram." The onboarding would kind of assume those plugins were already installed, and if they weren't, you would get weird errors later, and then you had to manually figure out what was missing. Now, it just installs them, which is the way it should have been from day one,
6:03but here we are. If you are setting up Open Claw for the first time and you're helping someone else set it up, that is a first run experience, and that is way smoother. The fourth major upgrade is {slash} models add. You can now add a provider model directly from inside the chat. Before, if you wanted to add a new model, you had to go edit the config file, you had to restart the gateway, and you had to make sure everything reloaded properly. Now, you just type in {slash} models add, and that's it. The provider name, the model ID, and the everything just works seamlessly. It's registered, you can use it immediately, no restart required even. So, thank you for Tac Hoffman for that one. Praise Jesus. This is the kind of quality of life thing that you do not appreciate
6:49until you really need it, and then once you have it, there's no going back. The fifth thing, and this is one for the enterprise folks watching, Claude Opus 4.7 on the Amazon Bedrock mantle is now callable through proper bearer auth route. Before, you had to hack it in, like AWS bearer tokens were being treated like Anthropic API keys, which is not how it's supposed to work. Now, it's done right. So, if you are on a team that has to use Bedrock for compliance reasons, or you have AWS credits you need to burn, you could finally call Opus 4.7 through the mantle the way it was meant to be called. Shoutout to Weir Joe on that PR. So, there is also a bunch of other provider stuff worth mentioning quickly. Tencent
7:34Cloud is now bundled. You can now onboard through Token Hub. They have the Hay 3 preview models in there with tiered pricing, OpenAI responses now use OpenAI native web search tool automatically when web search is on and you have not pinned a specific search provider. GPT-5 prompt overlay moved into a shared provider runtime. So now if you're calling GPT-5 through open router or open code or Codex or any of those, you get the same behavior as if you called it directly, which is how it should have always worked. WhatsApp got some nice upgrades as well. There is now native reply quoting with reply to mode. So when your agent responds in WhatsApp, it can actually quote the original
8:19message like a real person would. If you're struggling to keep up with content, well, I'm about to save you about 40 days worth of work. I built something called content [music] machine. It's 10 AI agents that run on the open claw orchestration and they handle everything. Scripts, thumbnails, X posts, [music] 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 [music] any niche, fitness, finance, real estate, marketing, whatever you are building and it is 100% completely customizable to
9:07your use case. >> [music] >> 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. [music] $97 one time, it's not a subscription. I'll put the link down below and you'll thank me later. Thanks to MC Axter for that one. And for WhatsApp groups, the direct messages per chat system prompts now get forwarded to inbound contacts. So, if you have a different behavior configured for different groups, that behavior actually fires on every turn now. But before, it was kind of a coin flip. You never knew what you were going to get.
9:52Shout out to Blue T Go for that one. And the sessions and control UE got some love, too. There is a mailbox style filter on sessions list now. So, you can essentially filter by label and by agent and by search, and the title previews work properly. It feels like an inbox. The control UE also added browser local personal identity for the operator, so you can set your name and avatar, and it quickly shows up in the chat. Quick settings got tightened up, as well. Agent fallback chips are cleaner. Narrow screen chat layouts were improved. Thanks to Burns Dev and Dango Zeng for that all those beautiful upgrades. Gateway diagnostics are now way better. There's a payload free stability recording that is on by default, and now
10:38you can export support ready diagnostic bundles that include sanitized logs, and status health, config, and stability snapshots. So, if you're ever debugging an issue with somebody on Open Claw or their team, you can just ship them in this bundle, and they can see what's going on. Shout out to Goomadeers for that one. There is also some key fixes I want to mention. The default thinking level for reasoning capable models got raised from off or low to medium. So, if you have a reasoning model, and you have not explicitly set the thinking level, it now defaults to medium, which is nice. Which means it's smarter responses out of the box. Plugin install now adds newly installed plugin IDs to your plugins the allow list before enabling.
11:24So, you do not get weird permission errors. Memory search switched to SQLite vec to KNN for vector recall, while keeping the full post filter results within its limits, which means semantic search is faster and more accurate in multi-modal access. So, there's a fix for WhatsApp inbound and outbound sends where a reconnect drain was duplicating cron sends to 7 to 12 times after the 30-minute inbound silence watchdog fired mid-delivery. If you have ever had open claw spam the same message a bunch of times in WhatsApp, that's why. This is fixed now. So, there is also a Linux fix where gateway wraps its child processes into a shim that raises child own score ADJ. So, if the kernel runs out, if it
12:11runs out of memory, it kills the transient workers first instead of taking down the long-lived gateway, which is the right behavior. Status got an explicit runner field now, so you can basically see whether a session is running on in-bed Pi or CLI back-end provider or the ACP harness agent like codex or Gemini. So, the two things I want to just really mention here because these are the kind of things that you won't notice until you really actually use them. The first one is the plug-in load time is now 82 to 90% faster to native JD loading for built bundle plug-in modules, so the difference is like night and day. Doctor non-interactive runtime also dropped by about 74% because the doctor plug-in pass are now lazy loaded. So, open claw
12:56just feels snappier across the board. The second thing, token juice is now bundled natively in open claw as an opt-in plug-in. It compacts noisy exec in bash tools in Pi embedded runs, which means your token bill goes down and your context window stays cleaner. Shout out to Vincent on that one for the JD work. So, 4.22 is pretty insane. Like, there's a ton of features in here that we could all use that will dramatically change our workflows. Like, if you were to summon up, voice agents just became usable across every provider, setup just got way easier, and you can add models from the chat without restarting the gateway. And enterprise users can finally call Opus 4 7 through bedrock the right way. Plus, a bunch of speed and stability stuff under the hood. And I love this release. I think it's one of
13:42the best in the last week, right? So, if you're not already running Open Claw, go and get it. The link is in the description, and if you want to learn how to actually build AI agents with the stuff, come join us in the community. I launched the community 31 days ago. We have 214 members, and we do nine live calls every week. We just added a local LLM call where you're going to run and learn how to install local models to run your workflows. So, we're going to be doing that call on Sunday, and then we do eight other calls during the week. So, come join, and 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 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.
Claude Code Just Dropped a Massive Update. Voice Mode, Loop Commands, and a Million Token Brain.
Anthropic shipped a dozen Claude Code updates in March. Voice mode, loop commands, a million token context window. Here is what actually changed my workflow.
OpenClaw 5.4 Update: Google Meet Voice Agents and Reliability Fixes
OpenClaw 5.4 (2026.5.4) is mostly about voice and reliability. Google Meet and Twilio calls can now run through a realtime Gemini voice bridge with paced audio and proper barge-in handling. The release also fixes problems on Windows and Docker, in Codex transcription, plugin loading, channels, cron diagnostics, command safety, and memory.
