OpenClaw 4.27 - Desktop Control With Codex

追加: 2026-05-07

[00:00:00]In this video, I'm going to be talking about the newest OpenClaw update, which has some amazing features introducing Codex Personal Computer, which just dropped about an hour ago. And the headline feature is a pretty underrated thing. And like I said, Codex computer use has been making its way all around X lately as far as the capabilities, being able to put build out agentic workflows, and they put that directly inside of Open. So, it can now actually take control of your desktop, mouse, keyboard, screen, the whole thing. You know, the first thing you need to understand and know is that now OpenClaw ships with a native setup flow for Codex computer use. There are two new commands. So, there's a {slash} Codex computer use status command, which basically checks whether the MCP server you need is installed and reachable. And the {slash} Codex computer use install command, which actually goes out and finds the right marketplace entry and installs it for you. So, there is also an optional auto-install setting for people who want zero friction. And the most important part is that before any Codex mode turn starts, OpenClaw runs a fail-closed MCP server right back and checks to see if it's good to go. So, if the desktop control bridge is not healthy, the turn does not run. It refuses to send a click or a keystroke into the void, which is generally pretty cool, you know, because computer use has been one of those features that everyone talks about and nobody actually uses.

[00:01:35]So, setting it up was a pain. You had to know which MCP server to install, and then you had to wire it up in the right path. You had to manually install the peekaboo or the CUA driver or whatever bridge you wanted. Half the time it would silently break and you would not know until your agent just stopped responding mid-task. Now, it's just one command to check and then one command to install. And a safety net that catches the bad cases before they can cause damage. For non-technical people, this is the kind of feature you have been really waiting for, right? Imagine you're running an OpenClaw agent and you tell it to open Photoshop, edit a thumbnail, save it, and upload it to your mission control dashboard. Before today, that was like three different tools, like a custom script and probably a YouTube tutorial that did not match your version. Now, it's just an agent that controls your computer, and that is a different category of capability entirely. The docs even spell out how Codex computer use, the direct CUA driver MCP, and OpenClaw app peekaboo bridge fit together, so you can pick the path that makes sense for you and your machine. The second thing in this release is DeepInfra. They added DeepInfra as a bundled provider, which I know it sounds boring on the surface, but it's actually one of the bigger wins in 4.27. DeepInfra is one of those providers that runs a ton of open-source models at very competitive prices. So, if you're looking for a cheaper alternative, there you go. You got Llama, Mistral, Quinn. 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 Cloud Code 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.

[00:03:34]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 Cloud 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 who hold you accountable, and coaches who can help you when you get stuck.

[00:04:00]I'll put the link in the description down below. Get in now before the price goes up. The whole catalog. With OpenClaw 4.27, you set deepinfra_api and then put in your key in your environment variables, and you turn it on. And then you immediately get model discovery, image generation, and editing. Even image and audio understanding, text-to-speech, text-to-video, and memory embeddings.

[00:04:24]One API key and it covers all six of those capabilities. Provider-owned onboarding and provider-owned base URL policy, that's also included. So, if you have been running like five different providers for different media tasks, one for embeddings, one for image gen, one for text-to-speech, one for understanding, you can collapse a lot of that into DeepInfra and it's a lot cheaper. It's simpler to config, it's simple to use, there's fewer API keys that can get broken into rotate, and that's a great upgrade. That's a huge upgrade personally. Like I've been juggling different types of image generation models and 11 Labs for voice and OpenAI for embeddings. It's just a pain in the butt. So, now it's all in one. Now, you can just pick DeepInfra and then let it carry out a lot of that weight. You can still use the dedicated providers when you want premium quality, but for the volume work, this is a clean default. And because the catalog is OpenAI compatible and discovered dynamically, when DeepInfra adds new models, you get them automatically.

[00:05:26]There's no more waiting for a core release. The third thing is they did a really big plugin for metadata cleanup.

[00:05:34]This is more of a developer thing, but it actually matters for everyone because of what it does at startup times. So, they moved provider catalogs off of a runtime fallback hooks and into plugin manifests. So, there's Qwiiphan, Xiaomi, and Nvidia, Cerebras, Mistral, Moonshot, Deep Seek, Tens It, Token Step Fun, By Plus, Volcano Engine, Fireworks, Together AI, Shoots, Kilo, OpenAI, Open Goat, all of these provider catalogs now live plugged in manifests instead of being baked core. What does this mean for you? Faster gateway boot, cleaner provider rows in the model picker, and aliases that suppress when you can audit instead of a hidden runtime hook. So, if you are a plugin author, you can now declare your model rows as data instead of code. So, shipping a new provider does not require a core release. The platform gets faster, the catalog gets cleaner, and the rate of new provider support goes up. This is the kind of architectural cleanup that does not show up on a thumbnail, but it does set up the next year of growth. The fourth thing, and this one is for the local model crowd. They added local sandbox for Docker GPU passthrough. So, there is a new opt-in setting called sandbox.docker.gpu that lets you pass the GPU flag through to your sandbox agent containers. So, which this all means, if you're running a sandbox agent on a host with a GPU, you can now do local GPU inference inside of that sandbox. So, for someone like me, I'm running on a Mac Mini, this is mostly just a FYI, but if you are on a Linux box with a real Nvidia card and you want your sandbox agent to use it for embeddings, image generation, or local inference, you can now do that without breaking the sandbox boundary.

[00:07:24]That's a niche feature, but for the people who need it, it's an enormous upgrade. So, I like that. Some of the work I've seen people doing inside of our community, too, with local Quinn and local image models is exactly that kind of setup. So, if you're one of those people, this is release is right down your alley, right? The fifth thing, too.

[00:07:42]There's a lot of them, but this one here is operator-managed outbound proxy routing. This might be the favorite feature in this release if you're running anything in corporate or enterprise networks. They added a proper outbound proxy config. There's two settings: proxy.enabled and proxy.url.

[00:08:03]So, you can set OpenClaw It validates that the URL is a real HTTP forward proxy, and then it bypasses the loopback for the gateway itself, so your local control UE calls do not go through the proxy. And on shutdown, it cleans up the proxy environment and then the dispatcher state, so nothing leaks out.

[00:08:25]This sounds dry, but it's actually really big deal if you're trying to run OpenClaw inside of a company network has to go through a corporate proxy. This is what unblocks it. Up until now, it was kind of all hacked together. You had to set HTTP and proxy environment variables and hope every dependent every dependency respected them. Now, it's first-class configuration. It's all validated with proper cleanup on exit.

[00:08:48]So, if you have a coworker or a friend who is dying to run OpenClaw at work, but their security team blocks all the direct outbound, you can tell them about the new 4.27 update. The sixth thing is mobile node presence. They added authenticated node.presence.alive events for both iOS and Android paired nodes. What this means is when your phone goes to sleep or your app gets backgrounded, the gateway can still mark the device as recently alive without treating it as fully connected. So, your last seen timestamp stay accurate, your paired devices do not look like they have disappeared, and background wakes can keep the presence fresh. For people running the OpenClaw mobile app, a quality of life upgrade. You will feel it immediately. The annoying thing about most agent apps is that the moment you close them, they look offline forever.

[00:09:41]Now, they look like real devices that are that are recently active, even. So, they're paired, ready to go. Small change, but there's a big difference in how connected the whole thing actually feel. The seventh upgrade here in this is, if you're struggling to keep up with content, what I'm about to save you about 40 days worth of work.

[00:09:59]>> [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 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, [music] 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 [music] and join the community. So, you plug in your own thing and it molds it to you.

[00:10:55]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. The long tail of reliability. There is a really nice run of fixes in 4.27 that are going to make the day-to-day way smoother. Telegram outbound calls are now bounded so a slow Telegram send does not wedge the gateway replies. There is also an alias cash for bundled plugin lookups. So, WSL2 file scans do not slow you down. Slack got a 15-second pong timeout by default and a bunch of new socket mode knobs you can tune yourself.

[00:11:36]Slack media downloads got idle and total timeouts so a stuck file share cannot lock up unbound messages handling. So, gateway startup no longer blocks on primary model pre-warm which means Slack and other channels come online promptly even when provider discovery is slow.

[00:11:55]Windows restart handoffs now the supervisor properly and there is a new models.pricing.enabled flag. So, offline restricted network installs can skip the open router and the light LLM pricing fetches that used to fail noisily on the boot. So, the last one is actually pretty huge for anyone running open claw in an air-gapped or limited internet environment. Up until now, every cold boot would try to phone home for pricing data. If that failed, you would get an error noise even though your install was fine. Now, you can just turn it off and it's clean. The eighth and the last big one, I saved it for the last, is Tencent Wombo and QQ bot. They added Tencent Wombo external channel plugin to the official channel catalog with full docs.

[00:12:43]This is a guide for web socket bot DMs and group chats and they expanded QQ bot with full group chat support including history tracking, mention gating, activating models per group config and FIFO messaging queuing. So, C2C streaming with proper streaming controller life cycles and unified media upload with chunk support for large files. For most of my audience, this is not the headline at all, but for anyone trying to reach a Chinese audience, this is a huge upgrade. China has its own ecosystem. Wombo and QQ bot are the channels you actually need to be on. The fact that Open Claw now treats them as first class instead of an afterthought is the kind of thing that opens up the entire market to whole Eastern models and the Western ones never even reach it, right? So, if you are building for cross-border audiences, this matters a lot. If you zoom out and you look at 4.27, the pattern is really clear here, right? Codex computer use is the headline feature. Deep infra to collapse five APIs into one and plugin metadata cleanup so the platform gets faster and more extensible. Sandbox GPU passthrough for the local model crowd, outbound proxy routing for enterprise networks, and mobile presence so your phone shows up correctly. A long tail of reliability fixes across Telegram, Slack, and gateway startup. There's a Windows restart implementation and offline pricing and it expanded coverage for Tencent and QQ bot. So, this release does not have to have one single flashy thing the way that most past ones do, but what it does have is a bunch of things that taken together make Open Claw feel like an actual real piece of software. Every release that lands like this is a release that quietly takes more excuses off the table. Like you wanted desktop control, obviously, it's done in this update. You wanted one provider that does everything, well you got that. You wanted to run it at work behind a proxy, there you go. You wanted to run a long tail of stability across Telegram, Slack, and Windows, they did that for you. So, if you are in an older version, tell your agent to update itself. Then run the Open Claw doctor to make sure your configuration is clean.

[00:14:55]If you have been wanting to play with computer use, run {slash} codex computer use status to see where you stand and then {slash} codex computer use install if it is missing. If you are using deep infra anywhere in your stack, set the environment variable and turn it on inside Open Claw. You can probably retire two or three other accounts and save some money. And if you're still on the fence about Open Claw entirely, what are you thinking? This is the moment you need to jump in. The platform is hitting that mature stage where every release is sanding more rough edges than it adds new ones. That's a really good sign for a piece of infrastructure you're building on top of, especially if you're betting your entire business on it. So, if you're all stuck on any of this, come hang out with us in the community. We have 203 members that have joined over the last 30 days and we do nine live coaching calls a week and we also do unlimited one-on-one calls. So, if you want to book a million one-on-one calls, go ahead. We'll help you. The whole goal is just to help you take the next step in your AI journey. So, Open Claw 4.27, Codex computer use is the headline feature here. You know, deep infra is an added benefit and they keep shipping these updates like crazy. This is the fourth update in the last 7 days. So, if you haven't subscribed to the channel already, please do. I release videos like this every single day. We'll see you in the next one. Have a blessed day.