OpenAI Codex adds Record & Replay: demonstrate a workflow once, get a reusable skill

OpenAI shipped Record & Replay in the Codex desktop app on June 18, 2026 (app version 26.616). (Source: OpenAI Codex changelog) You demonstrate a workflow to Codex once — the official example is uploading a video with its metadata, thumbnail, and subtitles — and Codex turns that recording into a reusable “skill” it can run again on its own, without re-prompting.

Key facts:

• Record & Replay shipped in the Codex desktop app, version 26.616, on June 18, 2026.
• It turns one demonstrated workflow into a reusable skill Codex repeats without supervision.
• The feature is macOS-only and requires Computer Use to be enabled.
• It needs a paid ChatGPT account and is not available in the EU, UK, or Switzerland yet. (Computer Use itself reached the EU on June 16, 2026.)
• The same release adds thread hand-off between local and remote hosts and bulk actions for the Automations history.
• Separately, Codex CLI 0.141.0 (also June 18) adds end-to-end encrypted Noise relay channels for remote executors.

(Source: OpenAI Codex changelog; The Decoder, 2026-06-20)

What Record & Replay actually does

Codex could already automate from a written prompt. Record & Replay is different: it captures a concrete computer-use sequence — the clicks, fields, and uploads you perform — and saves it as a named skill. Instead of re-describing a multi-step task every time, you do it once and replay it.

OpenAI’s own example is publishing a video: you upload it with its metadata, thumbnail, and subtitles once, and Codex can then repeat that whole flow on demand. The feature lives in the desktop app’s Automations area, which is also where the new bulk actions — mark runs as read, archive history — now apply.

What this means if you’re building with Codex

The point is deterministic, repeatable automation at the GUI layer. For tasks with no clean API — legacy dashboards, internal admin tools, publishing flows — a recorded skill beats re-prompting an agent each time: fewer tokens burned, fewer wrong turns, and an output you can predict.

Treat recorded skills like code. Version them, review what each one is allowed to touch, and keep a human in the loop for anything irreversible. A skill that “uploads a video” is also a skill that can act on your account unsupervised — the convenience and the blast radius are the same thing.

The local-to-remote thread hand-off, plus the CLI’s encrypted Noise relay channels, point at distributed execution: start a task on your laptop and continue it on a build box. If you run Codex across machines or in CI, those encrypted executor channels are the part to wire up and audit first.

The catch

This is a desktop-app feature, not an everywhere feature. It is macOS only, needs a paid ChatGPT account, requires Computer Use to be switched on, and is blocked in the EU, UK, and Switzerland at launch.

Computer Use is also a real security surface. A skill records, then re-runs, real clicks on your machine — scope what it can reach before you let it run unattended. And recorded skills are brittle to UI change: if the page you recorded against moves a button, the replay can break or do the wrong thing. Re-test a skill after any tool it touches updates.

If you’re weighing Codex against the alternatives for unattended work, our Claude Code vs Codex head-to-head covers where each one wins, and OpenAI’s Codex on Amazon Bedrock covers running it inside your own cloud boundary. For the token-metering and fallback discipline that applies to any agent you let run on its own, see running Claude Code as a daily driver.

Sources

• Codex changelog — OpenAI, app v26.616 and CLI 0.141.0, 2026-06-18
• OpenAI’s Codex can now watch you work once and repeat the task forever — The Decoder, 2026-06-20
• openai/codex — Codex CLI source and release notes