Codex Computer Use

Codex Computer Use

The Codex CLI has no built-in browser or desktop control — no --browser, no --screenshot flag. Computer use is assembled from two primitives: an MCP server (Playwright) for driving a browser, and image input (-i) for verifying screenshots captured by external tools. The Codex app has an in-app browser and a Computer Use plugin, but those are app features, not CLI automation.

Command surface verified against codex-cli 0.142.5; on flag errors, recheck codex --help and codex exec --help — placement and availability shift between versions.

When to use

• Verify a change visually: rendered page, iOS simulator, desktop app screenshot
• Drive a local web app through a browser (click, navigate, read console) via Codex
• Compare before/after screenshots for regressions

Preflight

codex --version # confirm CLI present; stop and report if missing

Capability reality check

Wire Playwright MCP (one-time setup)

codex mcp add playwright -- npx -y @playwright/mcp@latest
codex mcp list --json          # confirm registered

Equivalent ~/.codex/config.toml:

[mcp_servers.playwright]
command = "npx"
args = ["-y", "@playwright/mcp@latest"]
startup_timeout_sec = 20
tool_timeout_sec = 120

Do the npx package download during setup (network-enabled shell) — a sandboxed Codex run may not be allowed to download it mid-task.

Drive the browser

codex -a never exec \
  -m gpt-5.6-sol \
  -s workspace-write \
  -c model_reasoning_effort='"medium"' \
  "Use the Playwright MCP server to open http://127.0.0.1:3000, exercise the signup flow, check the console for errors, and report only reproducible UI defects with selectors or routes." \
  </dev/null 2>/dev/null

• -s read-only suffices for observe-only browsing; use workspace-write when Codex should also edit source after what it sees.
• Global flags (-a, --search) go before exec; always end with </dev/null (stdin gotcha — open-but-empty stdin hangs the CLI).
• Final message → stdout; progress/thinking → stderr.

Verify by screenshot (macOS recipes)

Capture outside Codex, analyze with -i:

# Whole screen (terminal app needs Screen Recording permission)
screencapture -x /tmp/screen.png

# iOS simulator
xcrun simctl io booted screenshot /tmp/sim.png

# A URL, headless
npx playwright screenshot http://127.0.0.1:3000 /tmp/page.png

codex exec \
  -i /tmp/sim.png \
  "Inspect this iOS simulator screenshot. Identify layout overlap, clipped text, and visible error states. For each defect, name the likely source file to inspect." \
  </dev/null 2>/dev/null

Before/after comparison — pass both:

codex exec -i /tmp/before.png -i /tmp/after.png \
  "Compare these screenshots and identify visible regressions. Only report differences supported by the images." \
  </dev/null 2>/dev/null

The verify loop

For "make it look right" tasks, iterate:

1. Capture screenshot.
2. codex exec -i … resume --last with the new capture.
3. Apply or let Codex apply fixes.
4. Recapture.

Resume keeps prior context:

codex exec resume --last -i /tmp/after-fix.png \
  "Here is the page after your fix. Confirm the overlap is gone; list anything still broken." \
  </dev/null 2>/dev/null

Sandbox and network

MCP server traffic is the server's own process — the Playwright browser reaches the network regardless of the Codex command sandbox. The sandbox bounds what shell commands Codex runs.

Troubleshooting