mcpField
C
codex-computer-use

codex-computer-use

github.com/joelfranciscoTesting & quality095 installsUpdated 11w ago
Repository

About codex-computer-use

codex-computer-use is a skill that delegates UI and UX verification of an application to Codex. It supplies a browser-testing brief and reporting procedure, including limits when Codex falls back to jsdom.

Key features

  1. 01Prepare a self-contained verification brief
  2. 02Run Codex with browser network access
  3. 03Request checkpoint verdicts and findings
  4. 04Distinguish DOM checks from visual checks

Use cases

  1. 01Check an app flow against layout and behavior criteria
  2. 02Review browser output without importing raw screenshots
  3. 03Follow up when browser testing falls back to jsdom

Install

Copy the command and run it in a terminal.

$npx skills add https://github.com/joelfrancisco/skills --skill codex-computer-use

After installing, type/codex-computer-usein Claude Code to confirm it loaded.

Bundled filesagents/openai.yaml

Details

From the project README

Codex computer use — UI/UX verification delegate

Codex is markedly better and cheaper at computer use and UI/UX verification than the main loop (screenshots are token furnaces).

  1. Write a self-contained verification brief for Codex: how to launch the app (dev server command, port), the flow to exercise step by step, and what "correct" looks like (layout, copy, behavior, no console errors).
  2. Run — ONE Bash call that captures the output path and reads it back (an inline -o "$(mktemp)" loses the path): OUT=$(mktemp) && codex exec -s workspace-write -c sandbox_workspace_write.network_access=true -C <repo-root> -o "$OUT" "<brief>" </dev/null >"$OUT.log" 2>&1; echo "exit=$?"; cat "$OUT" Always close stdin with </dev/null (codex blocks on a non-TTY stdin pipe). network_access=true is required: the default sandbox blocks ALL network, including localhost, so codex cannot reach the dev server without it. Tell it to use its browser tooling, take its own screenshots, and end with PASS/FAIL per checkpoint plus anything unexpected it saw.
  3. Report only the verdict and findings — never pull raw screenshots into the main context.
  4. If codex reports the sandbox blocked the real browser launch and it fell back to jsdom: DOM behavior checks still count, but rendering/layout/visual claims do not — verify those yourself with the browser skill (Pinchtab) or claude-in-chrome instead.
  5. FAIL or ambiguous → inspect yourself (browser skill / claude-in-chrome) before fixing.

Comments

Markdown supported