OpenAI adds one-click Claude Code migration to Codex

• OpenAI shipped a one-click Codex migration flow that imports instructions, configuration, skills, MCP servers, hooks, subagents, and recent sessions, according to OpenAI's migration post and the new Migrate to Codex docs.
• The official mapping is unusually broad: settings.json becomes config.toml, slash commands become skills, and recent sessions from the last 30 days can become Codex threads and projects, per the migration guide and testingcatalog's settings screenshot.
• User reports say the app now exposes more of the agent harness directly, with visible subagents, an in-app browser, artifact and source panes, and CI status in chat, according to dejavucoder's subagent screenshot, dejavucoder's app report, and reach_vb's CI screenshot.
• The CLI update underneath this adds persisted /goal workflows, explicit permission profiles, plugin marketplace plumbing, and more explicit MultiAgentV2 controls, per the 0.128.0 release notes and LLMpsycho's release summary.
• OpenAI paired the migration push with a business claim: OpenAI's revenue post said API revenue is growing more than 2x faster than any prior model release, while Codex doubled revenue in under seven days.

You can read the official migration map, skim OpenAI's getting-started guide, and check the 0.128.0 changelog. The weirdly revealing bit is that the import flow reaches past settings into sessions and subagents, while testingcatalog's app teardown spotted hidden Remote Control and Connections settings before OpenAI formally announced them.

OpenAI is framing Codex as a switch, not a fresh install. In the official docs, the app checks both user-level files on disk and the open repo, lets users choose what to bring over, then re-checks after import and offers a follow-up migration thread for anything that still needs manual cleanup.

That follow-up step matters. The docs say Codex can directly migrate the parts it knows how to map, then open a new thread with the migrate-to-codex skill for leftover setup that does not land cleanly, including prompt templates, plugin or marketplace residue, and hooks or MCP configs that depend on custom auth or environment variables.

The product pitch is bluntly competitive. TheRealAdamG's repost cast the move as switching from Claude to Codex, while TheRealAdamG's Rob Jamaica repost described this as Codex's "Claude Code moment."

The migration guide lists a concrete source-to-destination table:

• Instruction files → AGENTS.md
• settings.json → config.toml
• Skills → Codex skills
• Recent sessions from the last 30 days → Codex threads and projects
• MCP server configuration → Codex MCP configuration
• Hooks → Codex hooks
• Slash commands → Codex skills
• Subagents → Codex agents

That is broader than the launch tweet implied. OpenAI's post mentioned settings, plugins, agents, and project configuration, but the docs also cover hooks, MCP servers, slash commands, and recent session history.

There are also a few Claude-shaped tells in the plumbing. A GitHub pull request in openai/codex, Consume ai-title from external sessions, adds support for Claude Code ai-title records during import, while the 0.128.0 release notes mention external-agent config import as part of the plugin workflow changes.

The app is getting more explicit about what the agent is touching. WesRoth's screenshot shows separate panes for progress, artifacts, sources, branch state, edited files, pull requests, and checks, which is a lot closer to a live workspace than a single chat box.

OpenAI's own Academy guide describes the core unit as a thread inside a project folder on your machine. The same guide says users work locally inside a designated folder, with selectable permissions, and recommends starting with the default model and modest privileges before escalating.

User reports filled in the rest of the surface area:

• dejavucoder's app report said Codex has an in-app browser where users can open pages and add comments for the agent.
• dejavucoder's follow-up said that mattered for frontend work and knowledge-base tasks because the app keeps those materials in one place.
• reach_vb's CI screenshot showed branch details and pending or passing checks directly in the chat workflow.

The CLI update under this UX push is doing real harness work. The 0.128.0 release notes added persisted /goal workflows with controls to create, pause, resume, and clear long-running tasks, which lines up with shyamalanadkat's screenshot showing a task in a "Pursuing goal" state more than an hour into execution.

The same release expanded permission profiles, added a codex update command, and made MultiAgentV2 settings more explicit, including thread caps and wait-time controls, per the release notes. LLMpsycho's release summary also called out session import, plugin marketplace support, and resume and interrupt fixes.

Visible subagents are the other notable shift. dejavucoder's screenshot showed Codex spawning two named explorer agents and exposing their trajectories. That lines up with a GitHub issue on per-agent MCP scoping, where a user described environments with six subagents inheriting the parent session's MCP servers by default.

Before OpenAI pushed the migration button, testingcatalog's app teardown spotted a batch of unannounced app changes in version 26.429.20946:

• A hidden Remote Control feature
• A new Connections settings section
• "Avatars" renamed to "Pets"
• A Keyboard settings category for shortcut management
• An onboarding widget for email, calendar, and file integrations
• Claimed 20% faster computer and browser use
• Upgraded slides and sheets support
• Annotation for browser, artifacts, and code

The same teardown also claimed browser use and computer use appeared disabled in the EU. testingcatalog's later screenshot shows "Enable Remote Control on this Computer" and "External migration" already present as beta toggles in settings, which suggests some of the hidden surface is already leaking into shipped builds.

The onboarding screenshots point in the same direction. embirico's screenshot showed Codex asking what kind of work the user does, from Engineering and Data science to Finance and Sales, with a toggle for personalized task suggestions. That is a much wider funnel than a coding-only terminal tool.