OpenAI Is Making It Easier to Move From Claude Code to Codex

Introduction

In AI coding tools, the hardest thing to replace is often not the subscription.

It is everything a developer has accumulated around the tool: a carefully tuned CLAUDE.md, dozens of small workflow tricks, MCP server permissions, slash commands, skills, project memory, and months of chat history.

Those files may sit quietly on a developer's machine, but together they become the operating manual for an AI coding workflow. Move from Claude Code to another tool and, without migration support, much of that setup has to be rebuilt from scratch.

OpenAI is now making that switch much easier.

According to OpenAI's current documentation, the ChatGPT desktop app can import supported setup and recent work from Claude Code, Claude Cowork, and Cursor. Codex CLI can import from Claude Code and Cursor, while preserving the existing setup rather than deleting or replacing it.

That sounds like a simple convenience feature. In practice, it changes how much friction developers face when moving between AI coding tools.

OpenAI Is Packing Up the Claude Code User's Setup

OpenAI recently consolidated its external-agent import instructions into a single workflow.

In the ChatGPT desktop app, users can open Settings → Import and select supported agents such as Claude Code, Claude Cowork, or Cursor. In Codex CLI, the corresponding entry point is:

/import

The current official documentation says the desktop app can import instructions, settings, skills, plugins, projects, and recent work. Codex CLI can import supported setup, project files, and recent chats from Claude Code or Cursor.

The important detail is that the import is a copy rather than a destructive migration.

OpenAI's documentation explicitly says importing does not change or delete the existing agent setup. In other words, your Claude Code or Cursor configuration remains where it was while Codex receives its own imported version.

That makes the process feel less like replacing one tool and more like bringing an existing development environment into a second one.

What Gets Moved Is More Than You Might Expect

The official import documentation provides a fairly detailed mapping of source configuration to Codex destinations.

The conversion is therefore not always a simple file copy.

A Claude Code instruction file such as CLAUDE.md becomes AGENTS.md. Claude Code's settings.json is translated into Codex's config.toml. Slash commands are mapped into skills, while subagents become Codex agents.

This distinction matters because a successful import does not necessarily mean that every item behaves identically afterward.

A name can change, the configuration format can change, and the execution model can change with it.

Projects Are Reused Rather Than Uploaded as a New Copy

Project migration is another useful detail.

OpenAI's documentation says existing project folders are imported as projects using the same folders. The workflow therefore does not require the entire repository to be duplicated simply to continue working with it in Codex.

The import screen also lets users choose what to bring over rather than forcing a single all-or-nothing migration.

That means settings, instructions, skills, plugins, projects, and chats can be reviewed as separate parts of the move.

Old Chats Can Be Continued After Import

There is another practical problem with migration: old conversations can be too large for the destination tool to process in one go.

Codex handles this through context management rather than simply rejecting the imported conversation.

OpenAI's current documentation says Codex CLI can import up to 50 chats from the last 30 days. The desktop import flow also supports recent chats and lets users continue working from imported projects or conversations.

The source article describes the experience as follows: when a migrated Claude Code conversation is too long, Codex can compact earlier content when the user first continues that imported conversation, freeing context for the next turn.

The practical goal is simple: the migration should not end with a giant imported transcript that immediately becomes unusable.

This Import Flow Has Been Expanding

The source article traces the feature through several iterations of Codex.

The article reports that /import first appeared in Codex CLI 0.140.0 with a narrower set of Claude Code imports. Later releases expanded the workflow to include additional migration areas, followed by Cursor skill support and synchronization for imported sessions.

The exact release history is useful for understanding how quickly the feature has evolved, but the current official documentation is the better reference for what is supported today. It confirms that the import workflow now covers a broader set of agent setup and recent work than the earliest versions did.

It Is Still a One-Way Move in Practice

The source article makes a sharper observation: the migration documentation focuses on importing into Codex rather than exporting Codex configuration back into Claude Code.

There is an important distinction here.

OpenAI's current documentation supports automatic updates from the original agent into the imported setup in the ChatGPT desktop app. Users can enable automatic updates from Settings → Import and review import history there.

That does not mean that edits made in Codex are automatically written back into Claude Code or Cursor.

The direction is still important: the original environment can continue supplying updates to the imported copy, but the imported copy is not presented as a bidirectional synchronization layer.

For developers, this means it is worth deciding which environment is the source of truth before turning on automatic updates.

OpenAI and Anthropic Are Approaching Migration Differently

The broader competition is not limited to OpenAI.

The source article points out that Anthropic has also worked on migration tooling aimed at bringing user context into Claude. Anthropic's current help documentation also provides an official data-export workflow for Claude users, including conversation history and account data.

That creates a meaningful difference in philosophy.

If a platform supports both import and export, users can move their accumulated context more freely. If migration mainly moves in one direction, the feature can also become a powerful way to reduce switching friction toward the destination platform.

The technical details differ, but the strategic question is the same: who owns the developer's accumulated AI workflow, and how portable is it?

OpenAI Also Put Codex Inside Claude Code

OpenAI had already moved in the opposite direction before expanding its import workflow.

Its open-source codex-plugin-cc project lets developers use Codex from inside Claude Code for code reviews and delegated tasks. The plugin includes commands such as /codex:review and /codex:transfer, with the latter creating a persistent Codex thread from a Claude Code session.

The repository is released under the Apache-2.0 license and is maintained by OpenAI.

The /codex:transfer workflow is particularly relevant to the migration story.

The plugin documentation says it can take the current Claude Code session and create a persistent Codex thread, allowing the work to continue directly in Codex. The transfer uses Codex's external-agent session importer and follows the same conversion rules used for Claude history imports.

That is a different experience from simply starting a new Codex conversation.

The goal is to carry the working context itself across the boundary.

Four Things Still Do Not Move Cleanly

A migration button can remove a lot of friction, but it cannot make two different coding agents identical.

The source article highlights four areas where developers should expect manual work.

1. Permissions

Claude Code can contain carefully tuned permission rules built around specific files, commands, tools, and workflows.

Codex has its own permission and sandbox model. An imported permission-related setup therefore needs to be reviewed rather than assumed to be equivalent.

OpenAI's documentation specifically recommends reviewing tool restrictions and permissions in imported skills and agents after migration.

The important point is that permission migration is about preserving the original intent, not simply renaming settings.

1. Hooks

Hooks are another area where behavior can diverge.

OpenAI's import documentation explicitly warns that imported hooks may behave differently after migration and should be reviewed.

A simple hook may transfer cleanly. A more complicated chain with conditions, environment assumptions, or asynchronous behavior may need to be redesigned for the destination tool.

So this is not always a conversion problem. Sometimes it is a workflow redesign problem.

1. The Model Changes Too

This is perhaps the most obvious difference, but also one of the easiest to overlook.

You can move your instructions, skills, MCP configuration, project files, and memories into Codex, but the model executing those instructions is still different.

A prompt or CLAUDE.md file that has been tuned for Claude's behavior may not produce exactly the same result when it is executed by an OpenAI model.

The configuration may have moved successfully while the behavior changes underneath it.

The source article notes that Codex is designed around OpenAI's own model and Responses-based workflow. Developers who want to route Codex through a custom provider need to treat that as a separate configuration task rather than assuming that an imported Claude model configuration will automatically carry over.

1. Chat History Has Boundaries

Not every conversation is eligible for migration.

OpenAI's current documentation says Codex CLI imports up to 50 chats from the last 30 days. The /import command also has operational limits: it is not available during a running task, in a remote session, or while connected to a local app-server daemon.

The current documentation also distinguishes between local setup and project-level setup, which means developers should check exactly where their files and configuration live before starting the import.

The result is a useful rule of thumb:

One-click import does not mean zero review.

The Real Asset Is Not the Model Subscription

Model benchmarks change quickly.

A model can lead one month and be replaced by another version soon afterward. What changes much more slowly is the developer's surrounding workflow.

CLAUDE.md, AGENTS.md, skills, MCP servers, permissions, project memory, commands, and accumulated conversations can become more valuable over time because they encode how a developer actually works.

That makes portability a competitive issue.

The more reusable these configurations become, the less dependent a developer is on one AI coding platform.

And that cuts both ways.

Open standards such as SKILL.md, AGENTS.md, and standardized MCP definitions make it easier to move workflows from one tool to another. The same interoperability that helps OpenAI attract Claude Code users can eventually make it easier for those users to leave Codex as well.

For developers, the practical strategy is straightforward:

Keep important workflow assets in portable formats whenever possible, and avoid making your entire development process depend on one proprietary tool.

常见问题

What can Codex import from Claude Code?

Codex can import supported instructions, settings, skills, plugins, project files, recent chats, MCP configuration, hooks, slash commands, and subagents. OpenAI's documentation also maps Claude Code project memories into Codex memories.

How do I import Claude Code into Codex CLI?

Start a local Codex CLI session and run /import. Then choose Claude Code, select the supported setup or project files and recent chats you want to bring over, and review the imported configuration.

Can Cursor be imported into Codex?

Yes. OpenAI's current documentation says Codex CLI can import from Cursor, while the ChatGPT desktop app can import from Claude Code, Claude Cowork, and Cursor.

Does importing into Codex delete my Claude Code configuration?

No. OpenAI explicitly states that importing does not change or delete the existing agent setup. The imported configuration is brought into the destination environment separately.

How many Claude Code chats can Codex import?

Codex CLI can import up to 50 chats from the last 30 days. The /import command also has restrictions on when it can be used, so it is best to run it from a normal local CLI session.

Will Claude Code skills behave exactly the same in Codex?

Not necessarily. Skills, hooks, permissions, plugins, and command templates can depend on the original tool's behavior, so OpenAI recommends reviewing imported setup before relying on it.

Can I transfer a Claude Code session directly into Codex?

OpenAI's codex-plugin-cc provides a /codex:transfer command that creates a persistent Codex thread from a Claude Code session. The plugin is designed specifically for continuing Claude Code work inside Codex.

相关工具

• OpenAI Codex: OpenAI's coding agent for software development workflows.
• Codex CLI: The command-line interface for working with Codex locally.
• Claude Code: Anthropic's coding agent and the main source environment discussed in this migration.
• Cursor: An AI-powered code editor supported by the Codex import workflow.
• OpenAI Codex Plugin for Claude Code: An official OpenAI plugin for using Codex from inside Claude Code.

Related Links

• OpenAI Import Documentation: Official guide for importing setup and recent work from other AI agents.
• Codex CLI Documentation: Official documentation for installing and using Codex CLI.
• OpenAI Codex Plugin for Claude Code: Official OpenAI repository for the Claude Code integration.
• Codex Plugin README: Installation and usage instructions for the Claude Code plugin.
• Claude Data Export: Anthropic's official guide to exporting Claude account and conversation data.
• OpenAI Developers: OpenAI's main developer documentation hub.
• OpenAI Devs on X: The OpenAI developer post referenced by the source article.

Summary

OpenAI's current import workflow removes a major barrier for developers who have accumulated months of configuration and context in Claude Code or Cursor. Instructions, skills, plugins, projects, MCP configuration, memories, and recent chats can now be moved into Codex through supported import flows.

The important limitation is that migration is not the same as perfect compatibility. Permissions, hooks, prompts, plugins, and model behavior can change after the move, so imported configurations should be reviewed before they become part of a production workflow.

The most portable AI coding setup is the one that keeps its valuable instructions, skills, and project knowledge independent of any single agent.