How to Use Codex Background Computer Use on Mac (2026)

On April 16, 2026, OpenAI shipped Codex Background Computer Use — letting Codex operate any Mac app while you keep working in others. This guide walks through enabling it, running your first parallel agent task, and using it safely.

Last verified: April 22, 2026

Before you start

Check you have:

• ✅ macOS 14 (Sonoma) or later — verify: Apple menu → About This Mac
• ✅ Apple Silicon Mac (M1/M2/M3/M4) — Intel Macs are not supported
• ✅ 16 GB RAM minimum (32 GB+ recommended for multiple parallel agents)
• ✅ ChatGPT Plus, Pro, or Business subscription
• ✅ Codex desktop app installed from chatgpt.com/download/codex
• ✅ Latest Codex version (Settings → About → Check for Updates)

Step 1 — Install the Computer Use plugin

1. Open the Codex desktop app
2. Click your profile icon (top right) → Settings
3. Navigate to Computer Use in the left sidebar
4. Click Install

This downloads a ~180 MB helper that runs the virtual display layer. Install usually takes 30–90 seconds.

Step 2 — Grant macOS permissions

When prompted, grant these two permissions in System Settings → Privacy & Security:

Screen Recording

Required so Codex can see what’s on screen.

• Go to Privacy & Security → Screen Recording
• Toggle Codex on
• macOS will warn about restarting — click Later, Codex handles it

Accessibility

Required so Codex can click and type with its own cursor.

• Go to Privacy & Security → Accessibility
• Toggle Codex on

Restart Codex when both permissions are set.

Step 3 — Run your first safe test

Start with a read-only task to see how it works, before anything that writes data.

1. Open a new Codex chat

2. Click the ⊕ button → Background agent

3. Prompt:

   “Take a screenshot of my Finder window and describe what folders I have.”

4. Codex spawns a virtual display session, opens Finder, takes a screenshot, and reports back.

You’ll see a small “Agent 1” pill in the menu bar showing the running agent. Click it for a live preview of what the agent is seeing.

Step 4 — Try a real cross-app task

Once you trust the read-only flow, try something with writes:

“Open the Figma file ‘Spring Campaign 2026’, export each frame as a 2x PNG to ~/Desktop/spring-campaign/, then open Finder to that folder.”

Codex will:

1. Spawn a virtual session
2. Open Figma
3. Navigate to the file
4. Run File → Export on each frame
5. Save to the specified path
6. Open Finder when done

You can keep typing in TextEdit / Notion / Slack while this runs — the agent has its own virtual cursor.

Step 5 — Run multiple agents in parallel

Click ⊕ → Background agent again while the first is still running.

Tier	Max concurrent background agents
ChatGPT Plus	2
ChatGPT Pro	10
Business	5 (per user)

Each agent gets its own virtual display. You’ll see them as pills in the menu bar (Agent 1, Agent 2, …). Click any pill to preview.

Step 6 — Use the in-app browser

The April 16 update added a full browser inside Codex. Useful for:

• Tasks that involve web research + desktop apps
• Logging into a site once and letting Codex use that session
• Downloading files directly into the agent’s working directory

Access: ⊕ → Browser or ask Codex to “open a browser tab to …”

Step 7 — Give the agent memory (optional)

To let Codex remember your preferences per project:

1. Settings → Memory → toggle on
2. Per-project: Right-click the project in the sidebar → Enable memory
3. Codex will now remember build commands, style preferences, failed approaches

You can view and edit memory in Settings → Memory → Manage.

Step 8 — Schedule wake-up tasks

For long tasks that should run overnight:

1. In the chat: /schedule → pick a time
2. Describe the task
3. Close your laptop

Codex uses a combination of its cloud environment + local app. When you wake your Mac, it resumes or presents the finished result.

Example:

/schedule 6am “Go through yesterday’s commits on the main repo, write release notes, post them as a draft in Notion, and summarize the riskiest change.”

Safety checklist

Before giving Codex broad permissions, decide:

• Which apps can it access? Restrict in Settings → Computer Use → Allowed apps
• Can it access 1Password / Keychain? Default: no. Don’t change this.
• Can it install software? Default: requires confirmation. Keep it that way.
• Can it use Finder beyond ~/Documents? Default: no. Set explicit folders.
• Can it access browsers with logged-in sessions? Default: only the in-app browser. Recommended: keep it that way.

Kill switch

At any time: ⌘ ⇧ K stops all background agents immediately and revokes their current permissions.

Audit log

Settings → Computer Use → Action Log shows every click, keystroke, and plugin call made by every agent. Review periodically.

Troubleshooting

Problem	Fix
”Permission denied” when agent tries to click	Accessibility permission missing — re-grant in System Settings
Agent sees black screen	Screen Recording missing, or another app is showing DRM content (Netflix, DRM-protected video)
Agent is slow	Background agents share GPU/CPU with your work; upgrade to 32 GB RAM or run fewer concurrent agents
Agent misclicks small UI elements	Use higher DPI display scaling or disable “Use more space” in System Settings → Displays
Plugin won’t install	Quit Codex, delete ~/Library/Application Support/Codex/plugins/computer-use, reopen Codex

What to try next

• Export + upload workflow: “Export last 10 screen recordings to MP4, upload to our S3 bucket, paste links in Slack #eng”
• Cross-app research: “Search LinkedIn for ‘VP Engineering Berlin’, save the top 20 profiles to a Notion database, draft outreach messages”
• QA automation: “Open our staging site in Safari, click through the onboarding flow, screenshot any errors, open a GitHub issue”

These three workflows highlight why background computer use is different from in-terminal agents: none of them can be done purely by Claude Code or Cursor.

Windows and Linux?

As of April 22, 2026:

• Windows: In internal testing, no public date
• Linux: Not on the public roadmap

If you’re not on macOS, the best alternatives today are:

• Claude Computer Use API — works headlessly on any OS (Docker container)
• Gemini Chrome agent — works in Chrome on any OS

Cost watch

Codex Background Computer Use is included with ChatGPT Plus ($20/mo) and Pro ($200/mo). However, long parallel tasks consume more tokens than typical chat. Heavy users (10+ hours of background work per day) should:

• Monitor usage in Settings → Usage
• Upgrade to Pro for higher caps
• Set per-project token budgets in Settings → Spending

Bottom line

In under 10 minutes you can have a Mac where multiple AI agents do work across Figma, Notion, your terminal, a browser, and any other app — while you keep working in parallel. Start with a read-only task, install the plugin, grant the two permissions, and expand trust gradually.

The April 16 Codex update is the first time “computer use” has shipped as a polished consumer product. Experiment with small workflows this week and you’ll find at least one that saves an hour per day.

Related

• What is Codex ‘for almost everything’?
• Codex Computer Use vs Claude vs Gemini
• Codex 2026 vs Claude Code
• Best AI browser agents (April 2026)