OpenAI has added Record & Replay to Codex, giving eligible macOS users a way to demonstrate a repeatable workflow once and turn that demonstration into a reusable Codex skill. The feature appeared in the Codex app changelog for version 26.616 on June 18 and was later listed in OpenAI’s ChatGPT release notes.
The pitch is straightforward: if a task is easier to show than to explain, Codex can watch the workflow, inspect the steps, draft a reusable skill, and use that skill later with the tools available in the current environment. OpenAI’s Record & Replay guide gives examples such as filing an expense report, booking a parking space, creating a configured issue, publishing a video, or downloading a recurring report.
That makes Record & Replay less like a normal prompt feature and more like a bridge between AI agents, desktop automation, and lightweight internal process documentation. Instead of asking a worker to write a perfect instruction set for a recurring task, Codex can observe the real path through the app and package the pattern into something reviewable and editable.
How Record & Replay Works
OpenAI’s guide describes the workflow as a recording session inside the Codex app. The user opens Plugins, chooses the option to record a skill, reviews the suggested prompt, gives Codex any helpful context, and then grants permission when Codex asks to record actions. The user performs the workflow on the Mac and stops recording when the task is complete.
After recording stops, Codex inspects what happened and drafts a skill. The generated skill is meant to explain when the workflow should be used, what inputs it needs, which steps to follow, and how to verify that the task succeeded. A user can then ask Codex to refine the skill, including details that may not be obvious from the recording, such as naming conventions, default fields, approval rules, or decision points.
On replay, the user starts a new thread and asks Codex to use the skill with the inputs that changed this time, such as a different file, issue, date range, destination, or customer record. Codex can then apply the skill using available tools including Computer Use, browser actions, installed plugins, or a combination of them.
The Useful Part Is Not Just Automation
The most interesting part of Record & Replay is not that an AI agent can repeat clicks. Plenty of automation tools have tried to record desktop actions before. The more useful shift is that Codex is turning the demonstration into a skill: a piece of agent-readable process knowledge that can be inspected, edited, reused, and improved.
That matters for work where the official process is scattered across screenshots, Slack explanations, onboarding calls, and muscle memory. A finance team may know how an expense tool should be filled out, a product manager may know exactly how a bug report should be configured, and a creator team may have a publishing checklist that lives in one person’s head. Record & Replay gives those workflows a path into Codex without making the user translate every screen, field, and exception into a long prompt first.
For developers, the feature also fits into Codex’s broader skill system. OpenAI’s skills documentation describes skills as reusable directories with instructions and optional scripts, references, or assets. Record & Replay gives users a demonstration-based way to draft one when the workflow is already known but awkward to describe.
The Limits Are Important
Record & Replay is not available to everyone. OpenAI says it is available on macOS for eligible users, requires Computer Use to be available and enabled, and initially excludes the European Economic Area, the United Kingdom, and Switzerland. For managed Codex deployments, disabling Computer Use also disables Record & Replay and related enablement flows.
It also works best on stable workflows with clear success criteria. A process that changes every week, depends on judgment that is not visible on screen, or requires a human to interpret unusual edge cases may still need a written operating procedure, a custom plugin, or direct integration work. OpenAI’s own guidance points users toward a separate plugin when they want to distribute a stable package across a team, bundle multiple skills, include app integrations, add MCP servers, or manage install metadata.
That distinction is useful. Record & Replay can help capture a workflow quickly, but it should not be treated as a replacement for proper product integrations when a process is business-critical, heavily audited, or shared across a large organization. A recorded workflow may be ideal for a stable internal task; a plugin or API-backed integration is still the cleaner answer for repeatable production automation at scale.
What Teams Should Check Before Using It
The security and privacy questions are not minor because recording means Codex observes the actions and window content needed to learn the workflow. OpenAI tells users to keep recordings focused on the task and avoid secrets or sensitive data. Its Codex plan documentation also notes that content processed through Codex, including screenshots taken by Computer Use, follows the user’s ChatGPT training data controls.
In practice, teams should record with test accounts or low-risk examples where possible, avoid entering passwords, API keys, customer records, payment details, or private employee data during a demonstration, and review the generated skill before relying on it. The skill should state the inputs it expects, the conditions where it should not run, and the checks Codex should perform before considering the task complete.
Administrators also need to decide where Computer Use belongs in their environment. If Codex can operate desktop apps, browser sessions, and plugins, then permissions, connected services, and account scope become part of the automation design. A useful Record & Replay rollout starts with narrow workflows, limited access, and a clear review habit rather than a blanket invitation to record every repetitive task.
Why It Matters
Record & Replay shows how AI work agents are moving beyond chat, code completion, and isolated browser tasks. The goal is not just to answer questions or generate scripts; it is to learn the way a person completes a real workflow and preserve that pattern for future use.
If the feature works well, it could lower the cost of turning everyday operational routines into reusable agent skills. That has obvious appeal for teams that spend time on repetitive tickets, reports, uploads, forms, admin tasks, and internal tooling that never quite earns a custom integration. It also raises the bar for governance because every captured workflow becomes a small piece of delegated operational authority.
The best near-term use cases are likely to be narrow and boring in a good way: stable workflows, predictable inputs, reversible actions, and clear completion checks. That is where a demonstrated process can become genuinely useful without asking Codex to improvise around sensitive systems or ambiguous business rules.
