| name | codex-computer-use |
| description | Ask the Codex CLI (GPT-5.5) to run local app verification that needs computer use — browser automation, simulators, screenshots, app launching, or independent runtime inspection. This is how GPT-5.5 is invoked for computer-use work. Use when the user asks Claude to test a flow, verify UI behavior, inspect a running app, capture screenshots, or report confirmation and feedback about implemented behavior that benefits from real computer use. Work machine only (requires an authenticated Codex CLI). |
| allowed-tools | ["Bash","Read"] |
Codex Computer Use
Use Codex (GPT-5.5) as a separate local verification agent when the task needs real
UI interaction, screenshots, simulator/browser/device state, or an independent
runtime check outside your current context. OpenAI's local computer use is currently
well ahead of what Claude can do natively — delegate to it rather than struggling
with SSH-unfriendly, screenshot-heavy verification yourself.
Do not use this for ordinary code reading, typechecking, linting, or tests you
can run directly. Launching apps, simulators, or browsers to verify the requested
work is fine without asking; ask first only if the run could disrupt the user's
environment beyond that (closing their apps, changing system settings, or acting on
real accounts or production data).
Preflight
command -v codex >/dev/null || { echo "codex CLI not installed — cannot delegate"; exit 1; }
[ -f "${CODEX_HOME:-$HOME/.codex}/auth.json" ] || [ -n "$CODEX_API_KEY$OPENAI_API_KEY" ] \
|| { echo "codex not authenticated — run 'codex login'"; exit 1; }
If either check fails, stop and tell the user; do not silently fall back to doing
fragile computer use yourself.
Workflow
-
Write a single, self-contained prompt. Codex won't wander or gold-plate, so
be literal: name the app/URL/simulator, the exact flow to exercise, what counts
as success, and what to report back. Ask it to save screenshots to a known
directory so you can read them afterward.
-
Run it with full local access (computer use needs to launch apps and capture
the screen):
SHOT=$(mktemp -d)
codex exec -s danger-full-access -C "$PWD" \
"Verify <flow> in <app/url>. Steps: <...>. Success = <...>. \
Save each screenshot as PNG into $SHOT and end with a short PASS/FAIL report \
naming exactly what you observed. Do not modify project files."
- Point Codex at a specific repo/app with
-C <dir>.
- Attach an input image (e.g. a failing screenshot the user pasted) with
-i <file>.
- For read-only inspection that must not touch anything, use
-s read-only
instead of danger-full-access.
-
Read the evidence, don't trust the summary. Read the screenshots Codex saved
(Read on the PNGs in $SHOT) and check its claims against them before you
report to the user. Codex is a verification agent, not an authority.
-
Report back what was actually observed — PASS/FAIL, the screenshots, and any
discrepancy between Codex's summary and what the images show.
Notes
- The user's
~/.codex/config.toml already runs approval_policy = "never" and a
full-access sandbox, so codex exec runs hands-off; the explicit -s flag above
keeps this skill correct even on a machine whose config differs.
- Override the model with
-m <model> only if the user asks; otherwise use the
Codex default.
- If a command comes back wrong once or twice, ask Codex for the correct invocation,
fix it here, and tell the user so the skill stays accurate.
