What if you could send a Telegram message and have your AI agent open a browser, fill out a form, and submit it, all without you touching the keyboard? That’s not a concept anymore. With OpenClaw’s latest release and Codex + Computer Use support, it’s something you can actually set up today.
I’ve been running OpenClaw for a while now, mostly for chat automation and personal productivity tasks. But the 2026.4.27 release added something that felt genuinely different: `/codex computer-use`. It’s the first time I’ve seen an AI agent go from “answering questions” to “actually doing things on my screen.” Here’s what it is, how to get it running, and what you can realistically use it for.
What Is Codex Computer Use?
Codex mode in OpenClaw is the integration with OpenAI’s Codex agent, a model built for task-oriented, multi-step work. It’s distinct from just using GPT-4 for Q&A. Codex is designed to reason through a task, take actions, and iterate.
Computer Use takes that a step further. Instead of just outputting text or calling APIs, the agent can literally control your computer, moving the cursor, clicking buttons, typing into fields, and reading what’s on screen. Think of it as giving your AI agent hands.
The bridge that makes this work on macOS is PeekabooBridge, part of the Peekaboo 3.0 ecosystem. Peekaboo is a macOS app that captures your screen and provides a controlled interface for AI agents to interact with it. PeekabooBridge connects OpenClaw’s Codex agent to Peekaboo, so when the agent needs to “click a button,” it routes that action through Peekaboo to your actual display.
The result: you send a message in Telegram (or any OpenClaw-connected channel), and your Mac starts doing things.
What You Need
Before you set this up, make sure you have:
- OpenClaw installed and running: any recent version, but 2026.4.27 or later for Computer Use support
- Codex mode enabled in your OpenClaw config (or enabled via slash command)
- Peekaboo 3.0.0-beta4 or later: macOS only; this is what provides the screen interface
- PeekabooBridge: installed as part of the Computer Use setup process
- A Mac running macOS (this feature is macOS-only for now)
- Accessibility permissions granted to Peekaboo in System Settings
One thing worth noting: Computer Use does require an OpenAI API key with Codex access. If you’re already using OpenClaw with GPT-4, you likely already have this, but double-check your API key has the necessary permissions.
How to Set It Up
The setup is simpler than you’d expect given what it does. Here’s the step-by-step:
1. Enable Codex Mode
If you haven’t already, enable Codex in OpenClaw. You can do this from any connected chat:
/codex enable
Or set it in your OpenClaw config file. Codex mode tells OpenClaw to route task-oriented requests through the Codex agent pipeline instead of the default conversational model.
2. Install Computer Use
With Codex enabled, run the Computer Use installer:
/codex computer-use install
3. Install and Launch Peekaboo
Download Peekaboo 3.0.0-beta4 from the Peekaboo project page. Launch it, and when prompted, grant it Accessibility access in System Settings → Privacy & Security → Accessibility.
/codex computer-use status
You should see something like:
Codex: enabled
Computer Use: active
PeekabooBridge: connected
Peekaboo: running (v3.0.0-beta4)
If PeekabooBridge shows as “disconnected,” make sure Peekaboo is actually running in the foreground and that you’ve granted all necessary permissions.
5. Test It
Try a simple test to confirm everything works end-to-end:
Use computer use to open Safari and go to example.com
The agent should ask for confirmation (depending on your trust settings), then proceed to open Safari, type the URL, and navigate to the page. If you see that happen, you’re good.
What You Can Actually Do With It
Once it’s working, the range of things you can automate is surprisingly broad. Here are five use cases I’ve found genuinely useful:
1. Auto-fill web forms
Got a recurring form you fill out, expense reports, intake forms, client onboarding? Describe the data once, and the agent navigates to the URL, fills in each field, and submits. This is especially useful for forms that don’t have APIs or Zapier integrations.
2. Control desktop apps
Not everything has a web interface. The agent can open any macOS app, navigate its menus, and interact with its UI. I’ve used this to batch-export files from apps that don’t support scripting, and to configure settings across multiple tools without touching each one manually.
3. Automate browser workflows
Logging into sites, scraping information from behind a login wall, clicking through multi-step wizards, the agent can handle all of it. It’s not a replacement for Playwright or Puppeteer for programmatic use, but for ad-hoc tasks? Much faster to just describe what you want.
4. Bridge apps with no API
This is where Computer Use earns its keep. If a tool doesn’t have an API, a webhook, or a Zapier connector, you’re usually stuck doing it manually. Computer Use sidesteps that entirely. The agent sees what you see and can operate any GUI the same way you would.
5. Handle repetitive GUI tasks
Resizing and exporting images, archiving old files in a GUI-only tool, clicking through approval queues, anything you find yourself doing the same way more than a few times a week is worth handing off to the agent.
Tips and Limitations
A few things to know before you go wild with this:
macOS only (for now). PeekabooBridge and Peekaboo are macOS-specific. Windows and Linux support isn’t in the current release. If that changes, it’ll likely come through a community-contributed bridge or a future Peekaboo port.
Approval settings matter. By default, OpenClaw will ask for confirmation before executing Computer Use actions. Don’t turn this off unless you’re very confident in what you’re asking the agent to do. An unintended click in the wrong place can have real consequences.
Don’t use it for unattended sensitive actions. Passwords, financial transactions, anything with irreversible effects, keep those behind a confirmation prompt, or better yet, don’t automate them with Computer Use at all. This is a tool for supervised automation, not a set-it-and-forget-it background process.
Speed is not the same as accuracy. The agent is good, but it’s not infallible. On complex or dynamic UIs (anything with a lot of JavaScript or unpredictable state), it may need a retry or two. Check the results of important tasks.
Screen resolution and scaling affect performance. If you’re running a non-standard display setup or scaled resolution, Peekaboo may have trouble accurately mapping screen coordinates. Standard retina displays work well; unusual setups may need tweaking.
Final Thoughts
Codex + Computer Use is the kind of feature that sounds gimmicky until you’ve actually used it to knock out a task that would have taken you ten minutes of clicking. Then it stops feeling like a novelty and starts feeling like something you’d miss if it went away.
It’s best suited for power users who already have OpenClaw set up and are comfortable with the idea of an AI agent having some level of access to their desktop. If you’re running OpenClaw as a personal assistant, handling messages, managing tasks, surfacing information, adding Computer Use is a natural next step.
It’s not magic. It’s a well-built bridge between a capable agent and your existing desktop environment. And right now, on macOS at least, it works well enough to build real workflows around.
If you set it up and find good use cases, I’d genuinely like to hear what you’re doing with it.
Not Ready to Self-Host? Try ClawTrust
Building agents locally is powerful, but managing infrastructure, security, and scaling takes effort. ClawTrust provides managed OpenClaw hosting — your agents run securely in the cloud with automatic scaling, no DevOps overhead, and 24/7 monitoring.
This post contains affiliate links. If you buy through them, I may earn a small commission at no extra cost to you.
About The Author
This is wild. I knew OpenClaw could handle text tasks but didn’t realize you could connect it to computer-use agents like this. Is there a risk of it doing something unexpected on your desktop if a prompt goes sideways?
Totally valid concern, Alex! Computer Use is powerful but you’re right to think about guardrails. A few things help: first, always run it in a sandboxed user account or a VM for anything non-trivial — that way even a bad prompt can’t touch your main files. Second, OpenClaw lets you set approval rules so actions above a certain risk level need a manual confirm before executing. Start with low-stakes tasks (opening apps, reading files) before letting it do anything write-heavy. The post covers this a bit but worth emphasizing — treat it like giving someone remote desktop access, because effectively that’s what it is.
Just set up OpenClaw last week and this is giving me a whole new set of ideas. Does Codex work on Windows too or is this a Mac-only thing?
Good news — Codex is cross-platform! It runs on Mac, Windows, and Linux. The setup steps are basically the same across OS. The Computer Use piece (screen control) also works on Windows, though some of the keyboard shortcut mappings differ slightly. One thing to watch on Windows: make sure you have proper accessibility permissions enabled for the agent, otherwise screen-reading features can be blocked by system security. Should be a smooth setup once that’s sorted.
The use cases here are insane. Automating repetitive desktop tasks with AI feels like the logical next step. I’ve been using Zapier/Make for web stuff but this goes way further.
Exactly right — this is the layer above web automation. Zapier and Make are great when there’s an API available, but Computer Use fills the gap for everything that doesn’t have one (legacy software, desktop-only apps, anything that needs a human-like interaction flow). The sweet spot is combining them: use Make/Zapier for the web/API side, and OpenClaw + Computer Use for the desktop side. Together you can automate pretty much any workflow end-to-end.