OpenAI Codex Can Now Watch You Work Once and Turn It Into a Reusable Automation

OpenAI shipped a genuinely useful addition to Codex on June 18, 2026: a macOS feature called Record & Replay that watches you complete a task, then turns what it observed into a reusable, editable automation you can hand off to Codex going forward. No prompting required. No scripting. Just do the thing, and Codex figures out the pattern.

How it actually works

You install the Record & Replay plugin for Codex, then simply perform your workflow on your Mac as you normally would. Codex watches, capturing the actions and window content it needs to understand what you’re doing. When you stop recording, it inspects the sequence and drafts what OpenAI calls a “skill.”

That skill gets saved as a SKILL.md file: a plain, readable, editable markdown document. It covers when to use the workflow, what inputs it needs, the steps to follow, and how to verify the result came out correctly. Because it’s just a file, you can open it, read it, and edit it if Codex got something wrong or if the workflow changes.

The replay part is straightforward. Start a new thread, tell Codex to use the skill, and give it whatever is different this time around. The file to upload. The date range for the report. The issue number to reference. Codex uses the skill as its operating instructions and handles the rest, drawing on Computer Use, browser actions, connected plugins, or a combination depending on what the workflow requires.

What makes this different from just describing a task in a prompt

The obvious question is: why not just write a detailed prompt? A few reasons.

First, some workflows are genuinely easier to demonstrate than to describe. If a task involves navigating a specific sequence of UI steps, clicking in a particular order, or handling edge cases that only become apparent when you’re actually doing it, showing Codex is faster and more accurate than writing it out.

Second, the output is persistent and structured. A prompt lives in one conversation. A SKILL.md file lives in your repository or shared folder, gets version controlled, and can be handed to the whole team. One person records the workflow; everyone else benefits from the automation.

Third, Codex is smart about when to use which skill. It keeps a summary list of available skills in context and loads the full instructions only when it decides a skill applies. To avoid crowding out the rest of your prompt, that index uses at most 2% of the model’s context window, or 8,000 characters when the context window size isn’t known. The skills exist in the background and Codex reaches for them when relevant.

What this means for teams

The team-sharing angle is where this gets more interesting for organisations. OpenAI explicitly frames skills as shareable assets. One employee records how they file expenses, publish a video, or book a recurring meeting room. That SKILL.md gets added to a shared location. Everyone else on the team now has access to the same automation, without needing to record it themselves or understand how to prompt for it.

If your organisation uses Codex with a requirements.toml configuration file, there’s a corresponding control: setting computer_use = false under [features] disables both Computer Use and Record & Replay across the board. Enterprise administrators who aren’t ready to roll out computer-control capabilities have a clean way to hold that back while still using the rest of Codex.

The skill vs. plugin distinction

It’s worth understanding where skills sit in Codex’s broader architecture. Skills are the authoring format for reusable workflows. Plugins are the distribution unit. If you want to share a skill with a colleague informally, a SKILL.md file is fine. If you want to package it properly, bundle multiple skills together, add MCP servers, include app integrations, or manage install metadata for a broader rollout, you’d package it as a plugin instead.

Record & Replay is a fast path to creating a skill. The more structured plugin format is for when you’re ready to treat that skill as a proper piece of software with a defined lifecycle.

A few practical things to know before you start

OpenAI’s documentation recommends telling Codex your goal and noting which inputs might vary between runs before you hit record. Use realistic inputs during the recording, but avoid anything sensitive: secrets, credentials, or personal data you wouldn’t want captured.

The geographic restrictions are worth noting if you’re outside the US. At launch, Record & Replay is unavailable in the European Economic Area, the United Kingdom, and Switzerland. Computer Use must also be enabled in your Codex setup before the feature becomes available, so check that first if the option doesn’t appear.

Windows and Linux users don’t have access yet. The feature is macOS-only for now, with broader availability expected later.

The bigger picture

Codex started as a coding assistant. By March 2026 it had more than two million weekly active users, and OpenAI has been building it steadily toward something closer to a general enterprise agent platform. Role-specific plugin bundles arrived in early June 2026, tailored for functions like sales and analytics. Record & Replay is the next step in that direction: capturing expert workflows from the people who actually do the work, rather than requiring someone to translate those workflows into code or detailed prompts.

The ambition is clear enough. If Codex can learn from demonstration rather than instruction, the barrier to automating a repetitive process drops significantly. You don’t need a developer to script it. You don’t need to describe it perfectly in words. You just do it once.

For anyone spending time on repetitive Mac-based workflows right now, that’s worth paying attention to.