computer

name	computer
model	claude-sonnet-5
effort	low
argument-hint	[--model <gpt-model>] [--effort <level>] <UI task to perform>
description	Delegate a desktop or browser UI task to a Codex subagent (GPT 5.6 Sol by default) using Computer Use, where it reads the Mac screen and drives apps by clicking, typing, and scrolling. Use when the user wants Codex — or "GPT-5.5" — to operate a GUI app such as Chrome, Slack, Finder, Xcode, Figma, System Settings, or a website form, whether they say "have Codex click/open/fill/drag…" or type `/codex:computer`. Do NOT use for code or file work (use codex:implement to change code, codex:review to review it); prefer Claude's own Chrome tools for browser-only tasks when available. Flag risky real-world actions (sending messages, purchases, entering credentials) before invoking.

Codex Computer Use

Delegate a desktop UI task to a Codex subagent running GPT 5.6 Sol (or another GPT model the user names) with Codex's Computer Use feature. The subagent will take screenshots and drive the Mac by clicking, typing, and scrolling.

Choosing the model

The -m flag selects the model. Default to gpt-5.6-sol, but honor any specific model the user asks for:

• Use gpt-5.6-sol unless the user names a different model.
• If the user specifies a model — e.g. "drive it with gpt-5.6-terra", "use the <name> model" — pass that exact string to -m instead. Don't validate or second-guess the name; Codex will error if it's unknown. Note the model must support Computer Use; if it doesn't, the run will fail and you should report that back.
• If they typed /codex:computer --model <name> <task> (or -m <name>), strip that flag from the prompt and use <name> as the model.

Choosing the effort

Reasoning effort is set with -c model_reasoning_effort="<level>". Default to low for Computer Use — each action is a screenshot + reasoning cycle, and low effort keeps the loop fast. Honor any level the user asks for:

• Use low unless the user names a different level.
• If the user asks in prose — e.g. "high effort" — substitute that level.
• If they typed /codex:computer --effort <level> <task>, strip that flag from the prompt and use <level> as the effort.

The command below shows -m gpt-5.6-sol and low effort; substitute the chosen model and effort.

Before invoking — confirm intent

Computer Use has real-world side effects (sending messages, making purchases, changing settings). Before spawning the subagent:

• Confirm the task is genuinely a GUI task, not something better done via CLI or code.
• If the task could touch sensitive scopes — sending messages, transmitting personal data, deleting cloud data, submitting forms, entering passwords, installing software, changing system settings — surface that risk to the user and get explicit approval before invoking. Because the run bypasses approvals and the sandbox (see below), Codex will not stop to confirm risky steps mid-run — Claude's up-front confirmation is the only gate, so build any "don't do X without confirming" constraint into the prompt itself.
• The Codex Computer Use app must be present at ~/.codex/computer-use/Codex Computer Use.app. If it's missing, tell the user to install it via the Codex desktop app (codex app).

How to invoke

Codex Computer Use runs as a spawned macOS helper app that the Codex agent drives. Invoke Codex with the feature enabled and approvals+sandbox bypassed — codex exec runs non-interactively with approval mode never, so any prompt-on-request path auto-denies, and Computer Use will fail with "approval denied" when it tries to attach to an app. Bypassing is required for Codex to drive the UI autonomously:

codex exec \
  -m gpt-5.6-sol \
  -c model_reasoning_effort="low" \
  --enable computer_use \
  --dangerously-bypass-approvals-and-sandbox \
  --skip-git-repo-check \
  -C "$PWD" \
  "<PROMPT>"

• -m gpt-5.6-sol — the model; default gpt-5.6-sol, or the model the user named (see Choosing the model). Must support Computer Use.
• -c model_reasoning_effort="low" — reasoning effort; default low for Computer Use (see Choosing the effort).
• --enable computer_use — explicitly turn on the Computer Use skill (default is on, but be explicit).
• --dangerously-bypass-approvals-and-sandbox — required in codex exec. Because exec's approval mode is never, Computer Use cannot ask for on-request approval; without this flag it auto-denies attaching to any app ("approval denied") and also hits a profile-access error. This flag lets Codex drive the UI autonomously (there is no -a on-request in this CLI). Since it grants fully-autonomous UI control with no per-action confirmation, confirm the user has authorized it before invoking, and hold back the run for genuinely risky tasks (see below).
• Always set an explicit Bash timeout. Computer Use runs are especially slow because each action is a screenshot + reasoning cycle, and the Bash default (120000 ms / 2 min) will kill the run after just a few clicks. Pass timeout: 600000 (10 min — the maximum the Bash tool allows) on every call. Multi-step UI tasks routinely exceed 10 minutes, so for anything beyond a couple of actions run the Bash call with run_in_background: true and poll its output — a foreground call cannot exceed the 600000 ms cap.

Writing the prompt

• Codex has no memory of this Claude conversation. State the goal, the starting app or URL, any credentials the user has already logged in to, and the definition of "done."
• Be explicit about constraints ("don't send the email until I confirm", "read-only — don't click any buttons", "only look at the currently-open window").
• If the task involves a specific app, name it exactly ("Google Chrome", "Slack", "Xcode 15").
• Codex will need the app in the foreground or launchable — mention if it needs to be opened first.

After the run

Summarize what Codex actually did (from its stdout), including any screenshots or confirmations. Highlight anything Codex confirmed with the user mid-run and anything it declined to do. Don't paste the full transcript.

When NOT to use this skill

• The task is in code/files, not UI → use codex:implement (implement) or codex:review.
• The task is browser-only and Claude has the mcp__claude-in-chrome__* tools → prefer driving Chrome directly from Claude; that keeps the user's approvals in one place.
• The task is something the user should do themselves (financial transactions, legally binding submissions, credential entry) → hand off to the user rather than automating.