OpenClaw 4.27: NEW Computer Use Agents! 🤯

Claw 4.27 just dropped and this one is about making everything more reliable. Deep Infra joins as a new built-in provider giving you cheaper access to dozens of AI models through one API key. Your agents can now control your desktop through codex computer use, file attachments actually work properly in chat now, and telegram, slack, and discord all got stability fixes that stop some of the most annoying bugs people have been dealing with. Let me break it all down.

Let's start with Deep Infra. This is a new built-in provider in OpenClaw 4.27. If you haven't heard of them, Deep Infra gives you access to a huge range of AI models. For example, open source ones like Llama, Mistrout, Quen, at some of the cheaper prices in the industry.

And the integration isn't just for chat, it can also cover image generation, image editing, image understanding, audio transcription, text-to-speech, text-to-video, and embeddings for memory search all through one API. Why does that matter? Because of course cost is an issue for a lot of people running AI agents and this might be a decent alternative. Plus it covers everything in one place.

So Deep Infra gives you a way to run a lot of this at a fraction of what you'd normally pay. You sign up, get a key, plug into OpenClaw, and your agents can use models that Deep Infra offers. So it could be for example image generation for social media posts, it could be transcriptions for customer calls, it could be video generation, etc. And the setup is pretty simple.

With the onboarding, Deep Infra shows up as an option, you enter your key, it discovers your available models automatically, and then it's done. Now here's the feature that caught my attention the most, which is Codex Computer Use is now set up inside OpenClaw 4.27. That means your AI agent can control your desktop, it can click buttons, it can open up apps, it can navigate screens, and it works through a setup flow. You run forward slash Codex Computer Use status to check if it's ready, forward slash Codex Computer Use install to set it up, it checks for the right MCP server, it can discover what's available through the marketplace, and it has fail closed safety checks.

Meaning if something isn't right, it stops instead of doing something crazy. Think about what this means. Your AI agent can now do things on your actual computer. So it can fill out forms in apps, it could navigate dashboards, it could move data around tools that don't talk to each other.

And for anyone running a business where you spend time on a repetitive computer task, this is a direct path to automating them. File attachments in chat got a real fix too. Before this update, if you tried to send a non-image file to your agent through web chat, like a PDF or a document, it would just get silently dropped, gone. No error, no explanation, the file just vanished.

OpenCore 4.27 now properly stages those files so your agent can actually read them. Images still work like before, but now PDFs, documents, spreadsheets, they all come through properly. If you're sending contracts, reports, or data files to your agent for analysis, this was a blocker. Now it should work.

Let me talk about the model selection improvements because this fixes a frustration that a lot of people had before. If your main AI model failed, maybe the API was down, maybe you hit a rate limit, OpenCore would sometimes quietly switch to a completely different model you didn't even ask for. Your agent would keep responding, but with a random backup model that might give you worse results and you wouldn't even know it happened. OpenCore 4.27 makes model selection more strict.

So if you pick a specific model, like for example, you set a LLAMA, LLAMA3 from the model picker, and then that model goes down, it fails visibly instead of secretly switching. So that means you know what happened. You can make a decision. If you do one automatic fallback, you explicitly opt in with a fallback list, which means no more surprise model switches.

And there's a related fix for cron jobs. So if you scheduled a task to run at a specific time with a specific model and that model wasn't available, the job would silently use whatever model was lying around. Now it fails closed. If your scheduled model isn't there, the job tells you instead of running on something random, which is much safer for anything business critical.

Now, if you want help picking the right models, setting up OpenCore, which ones to use, how to set this all up, all the updates, et cetera, you can get that inside the AI profit board. And we run four weekly coaching calls every week where members bring their OpenCore setups and we go through exactly stuff like this, how to configure models, how to set it up, how to use computer use, et cetera, how to set up AI agents like Hermes and OpenCore. We've got a 30 day roadmap in there as well to help you automate your business with it. And there's 2,900 business owners in that building with OpenCore and AI agents like Hermes too.

You can actually check out the member map, connect with people in your local area, link in the comments description, or go to the AIprofitboarding.com to get access. Now, let's get into the channel fixes because there are some real ones here. Number one, Telegram. If you've been running a Telegram agent through OpenCore, you might've hit a problem where your agent would stall on startup.

Invalid tokens would give a confusing error about delete web hook failing instead of just telling you the token was wrong. OpenCore 4.27 fixes that. So bad tokens now fail fast with a clear message. There was another Telegram issue too.

So slow outbound messages or slow file system scans on WSL2 would wedge your entire gateway. One slow Telegram send could freeze everything else. Now those calls are bounded with timeouts. So slow sends can't block the rest of your system anymore.

And generated images were actually getting dropped from Telegram replies. So your agent would create an image, send the text reply, and the image would just disappear. Images now get preserved and sent properly even after the text has already been delivered. Slack had its own set of problems too.

The web socket connection, which is how Slack talks to your OpenCore, would go stale. The old detection method relied on app level events to check if the connection was alive. But if nobody was messing your agent, there were no events. So the connection would break silently and your agent would stop responding.

OpenCore 4.27 uses proper ping pong timeouts. 15 seconds by default and you can tune it if you need to. Slack file downloads were also causing hangs. If someone shared a file on a channel with your agent and the download stored it would freeze your entire inbound message handling.

Now there are proper timeouts on file downloads. Stored downloads get skipped instead of blocking everything. And there's a Slack fix for reset phases. So if you set up a custom phase to reset your agent's conversation, for example like type in new session, that phrase was leaking into the new conversation.

So your fresh session would start with your reset phrase as the first message. That is now fixed. The trigger gets consumed properly. And discord had a threading fix.

So group and channel replies are now private by default unless your agent explicitly uses the message tool. Before always on agents in discord servers would automatically post replies to every message, which was noisy and unwanted. Now your agent can lurk and only respond when specifically called on. If you want the old behavior back, there's also a setting for that.

There's also a fix for long discord interactions. So if your agent takes a while to respond, maybe it's running a tool or doing compaction, discord's interaction timeout would fire and you'd get an error. OpenCloud 4.27 now handles those interactions asynchronously so slow responses don't trip timeouts. Mattermost had a weird bug where regular users were being processed twice in messages.

Once is a normal message and once is a system event. So your agent would see the same message twice and sometimes respond to both. That is fixed. Messages come through once now.

And the gateway startup also got faster and more reliable. Before OpenCloud would wait for your primary AR model to warm up before starting your chat channels. If your model provider was slow to respond, maybe for example OpenRouter was having a bad day, your telegram, your slack, and discord agents would sit there not working whilst the gateway waited. Now channels should start immediately.

Model warm-up happens in the background, your agents come online faster regardless of how long the model takes. And there's a proxy feature for businesses that need it too. So OpenCloud 4.27 adds operator managed outbound proxy routing. So if your company requires all internet traffic to go through a proxy server for security or for compliance for example, you can now configure that.

You can set proxy.enabled and proxy.proxyurl and all outbound traffic goes for your agent through the proxy that you set. It validates the proxy URL, bypasses loopback traffic, and it cleans it up properly on exit. For enterprise setups, this was a missing piece. The gateway got more resilient on Windows 2.

So restart handoffs, where OpenCloud needs to shut down and restart itself, used to sometimes fail on Windows and leave the system in a broken state. Now restart tokens get routed through the Windows supervisor properly. Failed restarts can be retried instead of getting stuck. Sessions got smarter about staying clean.

If a session was supposed to reset daily or after being idle, background tasks like heartbeat checks would accidentally keep it live. A session that should have rolled over at midnight would stay open because a 3am heartbeat counted as activity. That is fixed in the last couple updates and OpenCloud 4.27 builds on that. There's also a fix for clock skew.

So if time stamps get weird, they can't keep stale sessions alive anymore. And oversized session files no longer trigger automatic rotation backups on every write. That was causing performance problems for busy agents. The session store just stays as is now and there's a doctor command to clean things up if needed.

Memory got a reliability fix too. So the dreaming system, which organizes your agent's memories over time, was sometimes spawning too many background processes across multiple workspaces, like a chain reaction that would eat up resources and cause timeouts. Now there's a cap on how many dreaming processes can run at once and if your dreaming model isn't available, it retries once with your default model before giving up instead of just failing silently. And I know OpenCloud 4.27 might sound like a fix and polish release compared to some of the flashier updates, but this is the kind of update that matters the most, particularly when it's been so buggy lately.

One thing that I'll be careful of here is if you don't feel like you need to update, don't update. If you do want to update, you can just type it in the chat. But honestly, if you're happy with a setup and you don't feel like you need any of these updates or features, you don't need to update, right? That's the biggest thing here.

Nobody really tells you that, but that's the truth because sometimes updates can break up your existing setup and I would always create a backup before you update to the newest version so that you can restore later and that will save you some time too. And it's always good to keep a backup of your existing setup if you're really happy with it. So this is an improvement. Channels don't install, models don't secretly switch, sessions reset when they should, gateways that start fast, files that actually arrive, and also the computer use is super exciting too.

I'm excited to test that myself. These are the things that made the difference between an AI agent that you can trust and one that randomly breaks at the worst time. The contributor count is still massive, 198 people shipped code for the release and a lot of what they shipped is the unglamorous work that makes everything possible. Fixing edge cases, handling timeouts, making error messages useful.

Not that fun to do I can imagine, but that's what turns a cool project into something more reliable and hopefully this one is more stable than it's been recently. If you're ready to start building with OpenClaw or you've been running it and want to make your setup better, come join us in the AI Profit Boardroom. We've got full tutorials on how to set up this stuff, full courses on OpenClaw and Hermes. We've got daily tutorials, plus you can connect with me personally inside there too, and we have prompt libraries for both OpenClaw and Hermes.

And a member map so you can connect with other agent users near you. Link in the comments description or go to the AIProfitBoardroom.com