The AI programming war has set its sights on the belongings of rival users.
When switching AI programming tools, the most expensive thing is never the subscription fee, but the CLAUDE.md you've been editing for half a year, the dozens of tips you've accumulated, the authorizations you've clicked one by one on MCP servers, and the slash commands you can type with your eyes closed.
These things don't cost money to buy, but they are the instruction manual for Claude Code to do your work. Switch to Codex, and the manual becomes invalid; everything has to be written from scratch.
What truly makes developers reluctant to leave is precisely this collection of belongings that has been accumulated over half a year and stored on their own machines.
What OpenAI recently did was to package this entire collection and move it into its own home.
On August 11, OpenAI consolidated the instructions for importing external agents into a single document.
Open the ChatGPT desktop client, and in settings there's a new Import option that can scan for Claude Code, Claude Cowork, and Cursor; in Codex CLI, typing /import recognizes Claude Code and Cursor.
With a few clicks, the configurations, skills, plugins, projects, and chats from the last 30 days that these tools left on your machine are all moved over at once.
What's moved is a copy.
The official documentation is very clear: importing does not modify or delete your existing agent configurations.
It's equivalent to giving a free moving truck to the opponent's users, coming to your door to pack up, and leaving the room next door exactly as it was.
After going through all the documentation for this process, there's one word that never appears from beginning to end—/export.
You can only move in, not move out.
What Gets Moved Is More Than You Think
First, let's talk about what this truck can carry. The official list provides a "moving checklist comparison table":
Instruction files (mainly CLAUDE.md on the Claude Code side) are moved over and become AGENTS.md;
settings.json is moved over and becomes config.toml: Claude Code's mix of JSON and Markdown is unified into a single TOML file on the Codex side;
Skills remain Skills, plugins remain plugins, but whether plugins can be used directly depends on what follows.
MCP server configurations map to Codex's MCP configurations, Hooks map to Codex hooks, and subagents map to Codex agents.
The slash commands you wrote yourself: packaging a frequently used workflow into /xxx, shared by the team, after moving over they are no longer called commands; they are all converted into skills.
Commands that come with plugins are also converted on the spot during installation.
The name changes, the invocation method changes, and the scope of use may also change. So strictly speaking, this item is not a move but a modification.
Claude Code's project memories become Memories. Chats from the last 30 days go into ChatGPT chats.
As for "project migration," the official wording is that existing project folders become "projects using the same folder."
In plain terms, Codex just takes over the directory on your local machine; it does not upload or copy the entire repository to OpenAI's side.
The checklist is detailed down to instructions, settings, skills, plugins, projects, and chats, each selectable.
There is a detail here that best shows how seriously this work is done.
Old conversations moved from Claude Code are often too long for Codex to read in one go.
On other tools, this situation usually results in an error message, and you have to figure out the rest yourself.
Codex's approach is very thoughtful: when a conversation is imported, it conveniently keeps a record: how long this conversation is and how much quota it occupies.
When you return to this old conversation and send the first message after the move, it first automatically condenses the previous content to make room, then continues to answer.
You don't have to do anything, and you don't have to delete anything; you can just continue chatting after importing.
Codex's import feature didn't just appear out of nowhere today.
/import first appeared in CLI 0.140.0, when it could only migrate Claude Code's settings, skills, and chat history, and Cursor users had to copy them manually.
On June 9, Codex app 26.608 brought the import process to the desktop app and also included it in the first-run onboarding.
CLI 0.145.0 on July 21 was the real expansion, expanding to six migration areas: settings, MCP servers, plugins, sessions, commands, and project-level memory.
On August 7, 0.147.0 added Cursor skill import and subsequent synchronization of imported sessions.
Three months, three iterations.
It seems OpenAI can't wait to take over Claude Code users.
Only in, not out
In the entire documentation, there is no command to export Codex's configuration.
What you change in Codex will not be written back to Claude Code.
The desktop app does provide an auto-update toggle in settings, and you can check import history.
But note the direction of synchronization: its purpose is to continuously sync new things from Claude Code and Cursor to ChatGPT.
The previous provider keeps updating, and the new provider keeps receiving.
The CLI side has even fewer options.
Codex CLI 0.147.0 released on August 7 added two things: importing skills hosted by Cursor, and syncing subsequent changes to imported Claude and Cursor sessions without creating duplicates.
Which door you move in through determines how much connection you have left with your previous provider.
You think you've been configuring tools all along, but you've actually been accumulating assets.
Now, someone tells you these assets can be packed up and taken away, and they can keep updating with the previous provider.
Both companies have done this migration play, but in different ways.
In early March, Anthropic launched the memory migration tool Memory Import, targeting ChatGPT users, to import your preferences and context accumulated in ChatGPT and Gemini into Claude.
Its support documentation also states: you can also export Claude's memory for backup or migration.
Both directions are open.
On March 30, OpenAI released an Apache 2.0 open-source plugin codex-plugin-cc that lets Codex run inside Claude Code, turning its own code review and task hosting into slash commands in the competitor's terminal.
The plugin includes /codex:transfer, which converts the current Claude Code session into a persistent Codex thread. The plugin documentation states that it uses Codex's external agent session importer, the same conversion rules as today's import.
From "entering the competitor's home" to "taking people away." In over four months, OpenAI's actions have escalated.
Four things that can't be moved
Developers in the community who have done complete migrations generally say: 90% of configurations can be automatically format-converted.
The remaining 10% that gets stuck is what matters.
First, permissions.
The fine-grained whitelist you manually adjusted in Claude Code for six months clashes with Codex's sandbox, which only has three coarse levels: read-only, workspace writable, and fully open.
This isn't just a matter of renaming a file; you have to rethink the intent behind that whitelist and translate it into one of the three levels.
Codex only has three built-in permission configurations: read-only, workspace writable, and fully open.
Second, Hooks.
Claude Code's hook model is richer, with conditional grouping and asynchronous processing chains.
Codex has no equivalent.
Moving complex hook chains over is not conversion; it's redesign. The official documentation itself warns: imported Hooks may behave differently than before, so review them before use.
Third, the model itself.
The first two are luggage; this one is the person doing the work.
Codex natively only uses OpenAI's Responses protocol and doesn't recognize Anthropic models. That means you can move all your CLAUDE.md, skills, and MCP configurations into Codex, but from then on, it's GPT, not Claude, that takes over these configurations.
Those prompts were tuned to Claude's temperament, and now someone else is executing them.
Unless you set up a compatible gateway yourself, register it as a model provider, and create a separate profile pointing to it. In the entire migration checklist, this is the only item without a native solution.
Fourth, the boundary of chat history.
Only local sessions on this machine can be moved.
Conversations on the claude.ai website and Cursor web conversations cannot be transferred. The CLI also has a hard limit: the last 30 days, up to 50 messages.
Some developers counted Claude Code's skill fields: there are 18 in total, only 6 are universal, and the remaining 12 are only recognized by Claude Code itself.
Only one-third of the skills you write can be taken with you.
The official also provides a post-import checklist: permissions for skills and agents, MCP servers that need to be re-logged in, Hooks whose behavior may change, plugins that need to be manually followed up, and command templates with parameters and file paths.
One-click migration, then half an hour of manual checking.
That's the complete process.
What's contested is not users, but users' belongings
Model scores are something that changes every six months. Today you're at the top of the list, and next month the opponent's new version may push you down.
But configurations, memories, skills, and historical conversations don't change: they lie on developers' own machines, growing thicker day by day.
This belongings is both an asset for developers and a moat for OpenAI and Anthropic to retain users.
OpenAI's move is aimed at draining the opponent's moat.
But while draining others, its own is also getting shallower: to release water from others, you have to dig the channel first, but the channel is bidirectional.
SKILL.md, AGENTS.md, and standard MCP definitions are becoming de facto standards.
The more universal the format, the harder it is for anyone to lock anyone in.
What can be moved into Codex with one command today can theoretically be moved to another with one command tomorrow.
For developers, the approach is simple: bet your belongings on formats that can be taken away, not on a specific tool.
This article is from WeChat public account "新智元", author: ASI启示录
