codex-computer-use

name	codex-computer-use
description	`codex exec` 経由で Codex.app の Computer Use を呼び出し、macOS アプリやブラウザを読み取り・クリック・入力操作する。Claude Code など、シェルコマンドを実行できる他エージェントから GUI 操作を代行したい場合に使用する。Chrome/Safari/Finder などの GUI 状態確認、ブラウザ UI 検証、アプリ操作に使う。

Codex Computer Use

Purpose

Use the Codex.app bundled CLI as a Computer Use bridge for another coding agent. The goal is to let an agent that can run shell commands delegate GUI work to codex exec, instead of depending on native Computer Use support in that agent.

Use this pattern:

codex exec --sandbox read-only --ephemeral --skip-git-repo-check -C "$PWD" "PROMPT"

Install This Skill

Install from this repository:

npx skills add schroneko/codex-computer-use -g -a claude-code -a codex -y

Use the equivalent agent selection flags for other skill managers if they support Agent Skills.

Preconditions

Installing this skill is not enough by itself. The host machine also needs Codex.app's Computer Use plugin, the Codex.app bundled CLI, and per-app Computer Use approval.

Check these before diagnosing Computer Use failures:

which codex
codex --version
codex mcp list

When running from Claude Code, also confirm:

which claude
claude --version

Expected:

• codex should be the Codex.app bundled CLI. On macOS app installs, it is often available under /Applications/Codex.app/Contents/Resources/codex
• claude should run if the caller is Claude Code
• codex mcp list should include computer-use
• Target apps must be installed and running before get_app_state

If which codex points to a CLI that is not bundled with Codex.app and Computer Use fails with Apple Event authentication errors, prefer the Codex.app bundled CLI.

If a Homebrew-installed codex is shadowing the Codex.app bundled CLI, one possible fix is to remove the cask and put the bundled binary on PATH:

brew uninstall --cask codex
ln -s /Applications/Codex.app/Contents/Resources/codex /opt/homebrew/bin/codex

Only do this after confirming both conditions:

• The machine installed codex through Homebrew.
• The user wants the Codex.app bundled CLI to own codex.

First-Time Setup

1. Install and open Codex.app.
2. Confirm codex mcp list shows computer-use.
3. Confirm the target app exists and has a bundle identifier:

mdls -name kMDItemCFBundleIdentifier -raw /Applications/Google\ Chrome.app

4. Approve the app for Computer Use, either interactively or by editing the approval JSON after user consent.
5. Start the target app before running codex exec.
6. Run a read-only get_app_state smoke test before clicking or typing.

The app approval file usually lives under the user's home directory:

~/Library/Group Containers/2DC432GLL2.com.openai.sky.CUAService/Library/Application Support/Software/ComputerUseAppApprovals.json

Expected shape:

{
  "approvedBundleIdentifiers": [
    "com.google.Chrome",
    "com.apple.Safari",
    "com.apple.finder"
  ]
}

Agent Usage

Invoke Computer Use through codex exec from the calling agent's shell tool. Keep the prompt explicit and constrained. Claude Code can use this skill directly, but the same bridge pattern also works from any agent that can run codex exec.

Read-only app inspection:

codex exec --sandbox read-only --ephemeral --skip-git-repo-check -C "$PWD" \
  "Do not run shell commands. Use computer-use MCP only. Call get_app_state for Google Chrome exactly once. Do not click, type, scroll, drag, press keys, or switch apps. Report only whether the call succeeded or failed and the exact error if it failed."

Single click:

codex exec --sandbox read-only --ephemeral --skip-git-repo-check -C "$PWD" \
  "Do not run shell commands. Use computer-use MCP only. In Safari, inspect the visible page. Click the visible Increment button exactly once. Do not navigate away, do not interact with browser chrome controls, extensions, account UI, or other tabs. Report whether it succeeded."

Single input:

codex exec --sandbox read-only --ephemeral --skip-git-repo-check -C "$PWD" \
  "Do not run shell commands. Use computer-use MCP only. In Safari, inspect the visible page. Set the visible text input to 'hello-codex'. Do not submit forms, navigate away, or interact with other tabs. Report whether it succeeded."

Use --sandbox read-only to protect files. Remember that it does not make GUI clicks or typing read-only; the prompt must constrain GUI behavior.

Browser Choice

Prefer Chrome when it is installed and visible to LaunchServices:

ls -ld /Applications/Google\ Chrome.app
mdls -name kMDItemCFBundleIdentifier -raw /Applications/Google\ Chrome.app
open -a "Google Chrome"

Fallback to Safari when Chrome is missing, being reinstalled, or returns appNotFound:

open -a Safari

Browser Use is separate from Computer Use. If browser-use is not shown by codex mcp list, use Computer Use for Chrome/Safari GUI operations.

Approvals

codex exec cannot answer Computer Use approval elicitation interactively. If a target app is not approved, one of these errors can appear:

Computer Use approval denied via MCP elicitation for app '...'

Fix by either:

• Run interactive codex, request a harmless get_app_state for the target app, and choose Always allow
• Add the target bundle identifier to ComputerUseAppApprovals.json when the user has approved broad app access

Common bundle identifier examples:

com.apple.Safari
com.apple.finder
com.google.Chrome

Failure Handling

For Apple event error -10000: Sender process is not authenticated, verify that codex points to the Codex.app bundled CLI. Homebrew-only CLI can see tools but fail Apple Events.

For appNotFound("Google Chrome"), check whether Chrome is installed, currently being reinstalled, or not running. Start it and retry.

For sensitive pages, avoid quoting page text, URLs, filenames, account data, or UI tree details. Ask Codex to report only success/failure and exact errors.

Safe Prompt Rules

Always specify:

• Target app name
• Target page or window when known
• Allowed tool family: computer-use MCP only
• Forbidden actions: no shell commands, no navigation, no other tabs, no account UI
• Exact action count: click once, type one value, inspect once
• Reporting limits: no private page contents or UI tree dumps

Use a local throwaway page for destructive or uncertain tests before touching real pages.