Codex Record and Replay: macOS Skill Routing Guide 2026

Codex Record & Replay skill routing is the operational signal in OpenAI's June Codex update. The feature lets a macOS user demonstrate a workflow, then asks Codex to turn the captured pattern into a reusable skill. For AI engineering teams, the important change is not only convenience. A recorded UI workflow becomes automation context that can later drive Computer Use, browser actions, and plugins, which means routing policy now has to understand who captured the workflow, what permissions it needs, and how expensive each replay may become.

What happened in Codex Record & Replay skill routing

OpenAI's Codex changelog for June 18 lists Record & Replay in Codex app 26.616. The dedicated documentation says the feature is available on macOS, requires Computer Use to be available and enabled, and initially excludes the European Economic Area, the United Kingdom, and Switzerland.

The workflow is straightforward: a user opens Plugins in the Codex app, chooses Record a skill, reviews a suggested prompt, grants recording permission, performs the task on the Mac, and stops recording when the workflow is complete. Codex then inspects the captured actions and window content, drafts a skill, and describes when to use it, which inputs vary, which steps to follow, and how to verify success.

OpenAI positions it for tasks that are repetitive, preference-heavy, or easier to show than to describe: filing an expense, booking a parking space, creating a correctly configured issue, publishing a video, or downloading a recurring report. The docs also draw a boundary: if a team needs a stable distributable package, multiple skills, app integrations, MCP servers, or install metadata, the workflow should become a plugin rather than only a generated skill.

Why Codex Record & Replay skill routing matters for AI engineering teams

Recorded skills change the adoption curve for agent automation. Before this pattern, teams usually had to write runbooks, prompts, scripts, plugins, or MCP servers before an agent could repeat a business workflow. Record & Replay lets an operator transfer tacit UI knowledge into agent context by demonstration. That is powerful, but it also moves governance earlier in the lifecycle.

A replayable skill may include hidden preferences: field defaults, naming rules, approval paths, browser states, and app-specific timing. It may also touch private windows, customer identifiers, internal dashboards, or paid SaaS actions if the demonstration is sloppy. OpenAI's docs tell users to keep recordings short, use realistic inputs, avoid secrets and sensitive data, and refine the skill after recording. Platform teams should treat those as control requirements, not user tips.

The cost model also changes. A replay can combine Computer Use, browser actions, connected plugins, and model calls. If the same skill is reused across a team, the gateway needs to separate training-time capture, review-time refinement, and execution-time replay. Otherwise a single "Codex" line item hides whether spend came from one-off skill creation or repeated automation runs.

The router/operator angle for Codex Record & Replay skill routing

Codex Record & Replay skill routing should be operated as a workflow lane, not as generic chat traffic. The useful control plane split is:

1. Capture lane. Recording and skill drafting should require explicit user approval, short scope, and redaction checks before the skill is saved.
2. Review lane. Generated skills should be inspected for sensitive fields, implicit credentials, region restrictions, and destructive actions before team reuse.
3. Replay lane. Runtime executions should carry a stable skill ID, initiating user, target app, and cost budget.
4. Plugin escalation lane. When a skill becomes shared infrastructure, promote it to a governed plugin with versioning and install metadata.
5. Exception lane. Regional exclusions and Computer Use availability should be visible in policy, so routing can fail closed instead of silently falling back to a weaker path.

This is where an AI gateway becomes useful beyond provider choice. It should label requests by workflow role, preserve session IDs across browser and Computer Use calls, and expose audit records for capture, refinement, and replay. TheRouter readers can map this to the broader TheRouter AI routing documentation and the earlier analysis of Codex remote dispatch and enterprise routing.

What TheRouter users should watch or try

TheRouter users do not need to wait for every desktop agent to expose the same feature. The immediate exercise is to inventory existing "tribal knowledge" workflows that agents already perform: issue creation, release checklist updates, support report downloads, internal dashboard exports, and repeated browser publishing flows.

For each workflow, decide whether it belongs as a prompt-only skill, a recorded skill, a deterministic script, or a plugin. Recorded skills are best for stable UI paths with clear success criteria and low blast radius. Deterministic scripts or plugins are better for high-volume, team-wide, compliance-sensitive, or destructive actions. If your team is already evaluating multi-agent execution, compare this with the Codex multi-agent per-thread runtime routing analysis.

Decision checklist for Codex Record & Replay skill routing

Before allowing recorded workflow skills in production, ask:

1. Who is allowed to capture a skill, and who can approve it for reuse?
2. Does the recording path exclude secrets, customer data, and unrelated windows?
3. Can each replay be tied to a skill ID, user, app, model tier, and cost budget?
4. Which actions require human confirmation even during replay?
5. When does a recorded skill need to become a versioned plugin?
6. How will regional exclusions and Computer Use availability appear in routing policy?

Codex Record & Replay skill routing turns personal demonstration into reusable automation. That is exactly the moment operators need workflow metadata, permission boundaries, and cost attribution to move with the skill.