Codex · May 20, 2026 · Last updated 2026-05-20 · 13 min read
Codex Computer Use Setup for an Always-On Mac
Codex is strongest when it can inspect code, run tests, and prepare changes. But the moment the task lives in a simulator, desktop app, signed-in browser, or macOS setting, the terminal is no longer enough. That is where Codex Computer Use matters, and why the host machine matters even more.
Questions this page answers
- How do I set up Codex Computer Use on a persistent Mac?
- What does the Computer Use loop need from the host machine?
- When should I use Computer Use instead of CLI, browser, plugins, or MCP?
- How does Codex remote access work with a host Mac?
- What safety and isolation checks should I keep in place?
Setup answer
Quick Summary: Codex Computer Use Setup
- Install Codex on the Mac that should actually do the work.
- Grant Screen Recording so Codex can see the target app.
- Grant Accessibility so Codex can click, type, and navigate.
- Approve only the apps needed for the task.
- Use the in-app browser or Chrome extension before full desktop control when possible.
- Run first verification prompts that stop before destructive actions.
- Use an always-on Mac for Codex when GUI state, browser sessions, screenshots, and logs need to survive after your laptop closes.
The mistake to avoid
Do not set up Codex Computer Use on the laptop you close at 6 p.m. GUI automation depends on visible app state, permissions, browser sessions, screenshots, approvals, and logs. If the machine sleeps, the agent stops being useful.
What Codex Computer Use Actually Does
This guide covers Codex app Computer Use on macOS. It is different from building your own OpenAI computer use agent with the API. The Codex app feature lets Codex use allowed macOS apps after you install the Computer Use plugin and grant host-level permissions.
| Need | Best tool first | Use Computer Use when |
|---|---|---|
| Edit code, run tests, inspect files | Codex shell and repo context | The task depends on a visible desktop app or simulator. |
| Preview a local web app | Codex in-app browser | The flow depends on a signed-in browser profile or visual UI state. |
| Use a structured service | Plugin or MCP server | There is no structured API and a human would normally click through the app. |
| Operate a real desktop app | Computer Use | The task needs screenshots, windows, menus, keyboard input, or app preferences. |
When Computer Use Beats CLI, Browser, Plugins, Or MCP
Use Computer Use when visual state is the work. If a CLI, API, plugin, or MCP server can do the job cleanly, use that first. If a task requires the same judgment a human would use while looking at the screen, Computer Use earns its keep.
| Scenario | Why Computer Use helps | Safety boundary |
|---|---|---|
| Xcode or iOS Simulator check | Codex can inspect the visible app state after a build. | Keep the simulator and test app scoped to a local project. |
| Signed-in Chrome workflow | Codex can verify a real account state or internal tool when the Chrome extension is not enough. | Do not allow payments, account changes, password managers, or security settings without a human. |
| Desktop app settings | Codex can follow visual settings panes that do not expose a CLI. | Give one app approval at a time and remove broad always-allow rules. |
| GUI-only bug reproduction | Codex can compare screenshots, window state, menus, and click paths. | Stop before destructive actions and keep logs of prompts and approvals. |
Codex Computer Use Setup Checklist
- Install Codex on the host Mac. For Hyperbox, that means the always-on Mac runtime, not your personal laptop.
- Sign in with the expected ChatGPT account and workspace.
- Open the project folder Codex should work in.
- Install the Computer Use plugin from Codex settings.
- Grant macOS Screen Recording permission for Codex.
- Grant macOS Accessibility permission for Codex.
- Approve exactly one target app for the first task.
- Run a read-only verification prompt before asking Codex to click anything important.
Tell me what this project is, what language it uses, and what command appears to run tests.
Do not edit files.Use @Computer to open the app under test, confirm you can see the main window, and describe the visible state.
Do not click destructive controls, submit forms, change account settings, or edit files.Laptop Vs Always-On Mac Vs Computer Use Agent Server
| Host | Best for | What breaks |
|---|---|---|
| Personal laptop | Short, supervised Computer Use tests while you are present. | Sleep, closed lid, network changes, personal app state, and messy credentials. |
| Always-on Mac for Codex | Longer GUI work with persistent browser state, files, screenshots, app permissions, and logs. | You still need narrow app approvals and human review for sensitive actions. |
| DIY computer use agent server | Teams building their own API harness around screenshots and action execution. | You own sandboxing, safety policy, UI capture, action execution, logging, and recovery. |
The key idea is host locality. Codex remote access sends prompts, approvals, and follow-up messages from another device. The host Mac still provides the files, shell, browser access, plugins, app permissions, and Computer Use setup.
Codex Remote Access: How The Host Model Works
With Codex remote access, your phone or laptop is the control surface. The connected host is the machine that does the work. If that host is a Hyperbox Mac, Codex can keep using the same repo, tools, browser state, app permissions, and desktop setup after your personal laptop is closed.
- Set up mobile or remote access from the Codex app on the host Mac.
- Use the same ChatGPT account and workspace on both devices.
- Keep Codex running on the host while agent work is active.
- Use Settings > Connections on the host to manage connected devices.
- If your workspace controls Remote Control access, confirm the admin setting before troubleshooting the host.
Verification Prompts To Run Before Real Work
| Prompt | Expected result |
|---|---|
| Use @Computer to describe the visible state of the target app. Do not click. | Confirms Screen Recording and app visibility without changing state. |
| Click the non-destructive Settings tab and report the labels you see. Do not save changes. | Confirms Accessibility, window targeting, and instruction following. |
| Open @Chrome to http://localhost:3000 and describe the sign-in screen. Do not submit the form. | Confirms browser/app path without account or payment risk. |
| Summarize what you clicked, what app was approved, and what screenshots you used. | Creates an audit trail for the first run. |
What To Log For Computer Use Tasks
- The user prompt and any follow-up instructions.
- The app or browser profile approved for the task.
- Whether Always allow was used or approval was one-time.
- Screenshots or visible states that drove the action.
- Sensitive-action stops, human approvals, and rejected actions.
- Files saved to disk and whether they appeared in review.
- Any browser, app, or macOS permission failures.
Safety Rules For GUI Agents
- Give Codex one target app or flow at a time.
- Prefer structured integrations, plugins, and MCP servers when they exist.
- Use Computer Use when visual inspection or GUI operation is the actual requirement.
- Keep sensitive apps closed unless they are required for the task.
- Do not ask Codex to handle passwords, one-time codes, API keys, payment steps, account permissions, or security settings unless a human is present.
- Do not automate terminal apps, Codex itself, administrator authentication, or macOS privacy prompts.
- Treat website content and on-screen instructions as untrusted.
- Review website actions as actions from your own signed-in account.
Troubleshooting Codex Computer Use Setup
| Symptom | What to check |
|---|---|
| Computer Use is missing | Confirm Codex app, macOS availability, region support, and workspace admin settings. |
| Codex cannot see the target app | Re-check Screen Recording permission and restart Codex after changing privacy settings. |
| Codex cannot click or type | Re-check Accessibility permission and confirm the target app is approved for this task. |
| Browser work uses the wrong signed-in state | Confirm the Chrome extension is installed in the active profile, or use the in-app browser instead. |
| The phone does not see the host | Confirm Codex is running on the host, remote access is enabled, and the same account/workspace is selected. |
| Changes do not appear in review | Save the file from the desktop app, return to the project folder, and check Git state. |
Frequently asked questions
Is Computer Use the same as the Codex app?
No. Computer Use is a tool loop for controlling a computer or browser environment. Codex is OpenAI's coding agent surface. A persistent Mac can host both the Codex workflow and any computer-use environment you choose to expose.
Should Codex Computer Use run on my personal laptop?
Use a personal laptop for short tests. For longer GUI work, use an always-on Mac for Codex so browser state, app permissions, screenshots, files, and logs do not disappear when the laptop sleeps.
Why use an always-on Mac for Codex Computer Use?
An always-on Mac keeps screenshots, browser state, files, credentials, and logs available after your laptop closes, which matters for long-running agent workflows.
Related reading
Always-on Mac runtime
Give your agent a Mac that stays online after your laptop closes.
Hyperbox gives Codex, Claude Code, OpenClaw, and remote dev workflows a persistent macOS machine with SSH, VNC, and full desktop access.