OpenAI wants a head-to-head with Claude Code—using your own work

RUNTIMEWIRE INVESTIGATION — Original analysis

Original reporting by RuntimeWire, based on documents, reverse engineering.

Why it matters

Codex Replay would turn migration into an acquisition funnel: import successful Claude Code work, rerun it in Codex, and judge the switch on a task that already matters.

Reporting record

OpenAI is building a Codex feature that invites developers to rerun work imported from Claude Code and compare the result, turning the migration process into a product trial built around a user's own coding task.

RuntimeWire found the feature, called Codex Replay, in production build 8881 of OpenAI's Codex desktop client, inspected on September 14th. The bundled interface tells eligible users: "Replay a task you imported from another agent and compare the results."

The invitation is narrower than that wording suggests. Client-side eligibility logic requires a successfully imported session whose provider is identified as 'claude-code'. The user must also have completed desktop onboarding within the previous seven days, must not have already seen or dismissed the announcement, and must be enrolled in an enabled experiment.

Those conditions place Codex Replay directly after a developer has brought working history over from OpenAI's largest coding-agent rival. The timing gives OpenAI a chance to make its sales pitch with a task the user has already considered important enough to attempt and preserve.

From migration to comparison

OpenAI already provides an import flow for bringing instructions, settings, projects and recent work into Codex from other coding agents. OpenAI's public documentation says the desktop app can import from Claude Code, Claude Cowork and Cursor, while the Codex CLI supports Claude Code and Cursor. The CLI can import as many as 50 chats from the previous 30 days.

Codex Replay adds a competitive action after that transfer. Selecting "Replay a task" begins an installation flow for an OpenAI-curated plugin. Once the user accepts, the client reloads its plugin catalog and waits for an 'open_controller' tool exposed by a 'codex-replay' server. It then switches the user into Codex and opens a projectless conversation titled "Codex Replay."

The client also records whether the conversation's first turn finished and whether the controller tool was called. That instrumentation indicates OpenAI wants to measure whether invited users make it through the comparison flow, rather than merely counting impressions or plugin installations.

RuntimeWire's inspection establishes that the invitation, eligibility checks, plugin installation path and controller integration are present in build 8881. The archive was inspected statically and was not executed. RuntimeWire did not alter feature flags, install the plugin, test account endpoints or run an imported task through the comparison.

The evidence therefore does not establish that Codex Replay is generally available, that OpenAI has enabled the required server components, or that an invited user can complete the process today. It also does not show when the code first appeared because no earlier archive was available for a binary comparison.

OpenAI leaves the verdict undefined

The bundled code does not reveal how Codex Replay presents the two results or whether it attempts to score them. There is no evidence of automatic winner selection, exact restoration of the original repository state, model-level comparisons, cost measurements or timing analysis.

That distinction matters because reproducing an agentic coding task requires much more than preserving a prompt. Repository state, dependencies, environment variables, tool permissions and changes made during the original session can all affect the result. The inspected controller implementation was not sufficient to determine how OpenAI reconstructs those conditions.

OpenAI separately offers a feature named Record & Replay, which watches a demonstrated macOS workflow and packages it as a reusable skill. OpenAI's documentation describes that product as a way to repeat stable computer workflows through Computer Use, browser actions and installed plugins. Codex Replay's bundled onboarding language instead concerns imported coding-agent tasks and comparing results.

The naming overlap could change before release. The build identifies the new plugin and conversation as 'codex-replay', while the user-facing invitation shortens the action to "Replay a task."

A switching tool aimed at the moment of doubt

Coding-agent users routinely move between products when one agent stalls, mishandles a repository or produces a result they do not trust. Import tools reduce the cost of that movement by carrying instructions, projects and chat history into a competing client. Codex Replay would push one step further by asking the developer to evaluate Codex before the original task and context disappear from view.

That makes the Claude-specific eligibility check the central finding. OpenAI's public import system supports multiple outside agents, but build 8881 reserves this invitation for users with a successful Claude Code session import. The client code does not establish whether the plugin itself can replay work from Cursor or any other provider.

The tactic gives OpenAI a better conversion opportunity than a benchmark chart or canned demonstration. The user supplies the task, the repository and the standard for success. OpenAI supplies another run. Whether Codex earns the switch remains up to the developer, because the inspected build contains no evidence that OpenAI has decided how the winner should be measured.