Claude Code Betrayed Me? I'm Done—Codex Can Import It Directly

YunLab · Firsthand AI tooling

Claude Code was a tool I used constantly and kept recommending to friends around me.

I was the only person using the account, and the proxy IP I used did not change. Even so, the account was banned. I appealed, but the outcome did not change.

This is my own experience, not a claim that everyone will be treated the same way. It did change how I think about depending on tools like this.

If a tool turns on me like that, I'm done with it.

What I regretted most was not losing the application itself. It was losing the work I had done inside Claude Code, especially the accumulated conversation history.

Codex could still locate the related projects, files, and outputs, but that was not Claude Code's original conversation context. How a discussion unfolded, why a decision was made, what I tried, and what I abandoned—those threads had not truly come across.

I searched for a while and never found a complete way to preserve that original context.

Today, after updating the Codex desktop app, I suddenly noticed a new item in Settings: Import from another agent.

It can import supported settings, Skills, Plugins, Hooks, projects, and recent conversations directly from Claude Code. I ran it immediately.

About 40 seconds later, Codex recorded 54 successful imports and zero failures. All 25 selected Claude Code session threads appeared in Codex.

This was no longer just Codex finding my old project files. The past conversations came with them.

The result of my local test

The first files landed at 14:03:45 China Standard Time. The batch completed at 14:04:25, roughly 40 seconds later.

The 25 imported sessions were associated with three local projects. Their original dates remained visible in this batch, ranging from May 24 to July 31, 2026, instead of all being rewritten as newly created conversations.

The scope matters. Fifty-four successes mean that the 54 objects detected, selected, and processed in this batch were recorded successfully. It does not prove that my entire Claude Code environment was cloned without loss, or that every Plugin and Hook has passed a runtime test.

This batch contained no success or failure records for Settings, Instructions, MCP, Subagents, Slash commands, or Memories. Those categories may not have been selected, detected, or required. Their absence from this batch is not a failure result.

What Codex can import from Claude Code

OpenAI's official guide currently documents these main mappings:

Importing a project does not create a second copy of its files. Codex continues using the original folder.

Standard Claude Chat history is not supported by this flow. The supported source is recent local Claude Code work.

OpenAI first introduced the desktop entry point in Codex app 26.608 on June 9, 2026, then expanded the workflow in Codex CLI. The timeline is documented in the official changelog.

The complete desktop workflow

1. Open the ChatGPT / Codex desktop app.
2. Go to Settings → Import. If Import is not shown as its own section, open Settings → General → Import other agent setup.
3. Choose Claude Code. Codex scans the supported local setup and work it can detect.
4. Under Select items to import, choose from Tools & setup, Projects, and Chat sessions. You can import everything offered or handpick the items you want.
5. Select Continue and wait for the import to finish.
6. Open an imported project or session and inspect the result.
7. If the lower-left status card says a Plugin or connection still needs attention, select Finish and complete the authorization or setup.

If Codex already contains substantial customization, importing in smaller groups is safer than selecting everything at once.

How to use it in Codex CLI

Start a local interactive Codex CLI session:

codex

Then, inside Codex, enter:

/import

The command is /import inside the interactive interface. It is not a standalone codex import shell command. Choose Claude Code, then select the supported setup, project items, and recent sessions you want to migrate.

Codex CLI discovers up to 50 sessions from the last 30 days. That 50-session maximum is documented specifically for CLI and should not be treated as a universal desktop limit.

/import is unavailable while a task is running, in a remote session, or while the CLI is connected to a local app-server daemon.

See OpenAI's /import documentation for the current CLI instructions.

How I verified more than a success screen

I did not stop at the completion message.

First, the local import history contained one clear Claude Code batch: 54 successful objects and an empty failure array.

Second, all 13 Skill directories were created during the import window, each with a SKILL.md file that Codex could detect.

Third, all 25 selected sessions had their own Codex task records and local session files. None was archived, and the earlier dates remained visible.

Fourth, the imported elements-of-style Plugin was callable in the current Codex task. I used it while tightening and organizing this article.

That proves this specific imported Plugin capability worked. It does not replace individual testing of every other imported Plugin.

What to review after importing

1. Open two or three recent sessions and check their titles, dates, and core context.
2. Run one safe Skill or Plugin in a real task.
3. Read imported Hook commands before allowing them to execute.
4. Check whether Plugins or MCP connections require a fresh sign-in or authorization.
5. If you imported Instructions or Settings, inspect AGENTS.md and the Codex configuration before relying on them.

“Installed” does not prove that a Plugin's external account is authorized. “Imported” does not prove that a Hook has been trusted to run.

The current public user guide does not document a one-click rollback workflow. If Codex already has important customization, a selective import is more prudent than importing everything in one pass.

My conclusion

The real value of this feature is not lower installation cost. It is lower continuity cost when changing tools.

Moving from Claude Code to Codex used to mean rebuilding Instructions, Skills, Plugins, MCP connections, and historical context by hand. Codex can now recover a meaningful part of that working environment.

This is not a perfect, invisible clone. Formats are converted, Plugins may need new authorization, and Hooks may behave differently. The correct workflow is:

Import, inspect, run one real task, then rely on it.

In this test, all 25 selected session threads were visible, 13 Skills physically landed, and one imported Plugin took part in writing this article.

That closes the part I regretted most after losing access to Claude Code.