> 📥 The full guide is available as a PDF (7 pages, Danish). Download it at the top of the page.
I moved my entire Claude setup into ChatGPT in under ten minutes.
Projects, skills and recent chats were detected automatically. My existing folders stayed exactly where they had always been. Afterwards a few connections had to be re-checked and re-approved, but the move itself was surprisingly simple.
That is worth pausing on. I have spent months building projects, instructions, skills and quality checks in Claude Cowork. By now that setup matters as much as the model. It tells the AI colleague how I write, where my files live, and when a deliverable is good enough.
I have argued before that Claude's brain lives in files you own yourself. I thought portability was mostly an architecture point. OpenAI has now made it practical with an official import function for Claude Cowork, Claude Code and Cursor.
> Key number. Under ten minutes — that is how long my own import of projects, skills and recent chats from Claude to ChatGPT took.
1. OpenAI built the moving truck itself
The feature is called "Import from another agent" and lives in the ChatGPT desktop app. OpenAI's official documentation covers the steps, the exact translation of file types, and the follow-up work you need to finish afterwards.
The desktop app can import from Claude Code, Claude Cowork and Cursor. Codex CLI can import from Claude Code and Cursor, but not from Cowork. The import can bring instructions, settings, skills, plugins, projects and recent work into ChatGPT or Codex.
The key word is "import". Your existing Claude setup is neither changed nor deleted, so you can test ChatGPT with the material you already work in while Claude keeps running as before.
That lowers a real switching barrier. Swapping the model takes seconds. It is the hours invested in folders, instructions, skills and habits that have made a platform choice hard to change.
2. How to drive the moving truck
1. Install or update the ChatGPT desktop app (Mac and Windows) with Chat, Work and Codex.
2. Open Settings > Import. If Import is not its own section yet, look under General for "Import other agent setup".
3. Pick your source. Only enable the sources you actually use.
4. Choose what comes along under "Select items to import": the whole setup, or Tools & setup, Projects and Chat sessions individually.
5. Start the import, then open an imported project or chat and continue from there.
6. Finish the setup. The status card lists plugins and connections that still need installation, login or approval.
7. Decide on sync. Automatic updates keep new and changed material flowing from the source.
3. What comes along
| From Claude or Cursor | In ChatGPT and Codex | What to know |
|---|---|---|
| Instruction files, e.g. CLAUDE.md | AGENTS.md | Read the imported file for overlap and vendor-specific wording. |
| settings.json | config.toml | Check permissions, models and local settings in particular. |
| Skills | Skills | Test both when they should trigger and when they should not. |
| Plugins | Plugins | Some plugins and underlying apps need manual follow-up. |
| Existing project folders | Projects with the same folders | Files stay put. Verify paths and where new deliverables are saved. |
| Project memories from Claude Code | Memories | The docs mention Claude Code specifically — do not assume all Cowork memory follows. |
| Chats from the last 30 days | ChatGPT chats | Older chats need a separate export if you want to keep them. |
| MCP server configuration | Codex MCP configuration | Custom logins, headers, env vars and transports may need new setup. |
| Hooks | Codex hooks | Behaviour can differ after import. Review before use. |
| Slash commands | Skills | Prompts with arguments, shell expressions or file placeholders must be test-run. |
| Subagents | Codex agents | Verify tools, permissions and responsibility boundaries. |
4. The best move happens in two stages
Stage one is a trial move. Pick one project, three to five core skills and a small set of recent chats. Then run three tasks you already know from Claude: a research task, a document task and a task involving an external connection. Compare quality, time spent and where files end up. Ask ChatGPT to describe the project's purpose, instructions, available skills and folders — if that description is wrong, the rest of the test is already at risk.
Stage two is scaling. Import the rest once the pilot works. Only enable automatic sync when you know both environments are allowed to process the same material.
5. Five things still need a human eye
1. Claude is untouched. Use the import for a parallel test rather than making the move final on day one.
2. History has a limit. Chats from the last 30 days only — archive older conversations separately if they matter.
3. Connections are not instructions. A skill can move; its access to Google Drive, Slack or an MCP server may not. Expect to log in again.
4. Skills must be test-run. One positive and one negative test each.
5. The import history is your checklist. Settings > Import shows history and missing setup.
6. The trip is not symmetrical
OpenAI documents a thorough path into ChatGPT. It does not document an equivalent export of the whole setup in the other direction. Anthropic offers official import and export of memory between Claude and other providers — but memory is only one part of the workplace. Projects, plugins, hooks, agents and platform state are something else.
> Principle. Keep your most important AI context in files you own. The model may change. The architecture should stay yours.
7. Never move customer data into the wrong account
The import makes the move technically easy. It does not decide whether the destination is the right one. If your Claude projects contain customer information, personal data, confidential documents or internal connections, import them only into the ChatGPT workspace and plan your company has approved. Check the data processing agreement, retention, access roles, admin settings and model-training settings first — and be careful with automatic updates.
8. Your next workhorse should win on real work
AI work is shifting from loose prompts to whole working systems: projects, instructions, skills, connections, memory and quality checks. When those parts are portable, the cost of trying a competitor drops sharply. That strengthens your negotiating position and lets you choose on quality, security and price.
Run the import with one project today. Pick a task you already know. Compare the result with Claude, then decide which workhorse deserves the next working week.
Sources
All links checked as of August 2026.
- OpenAI's official import guide: learn.chatgpt.com/docs/import
- ChatGPT Work and Codex: help.openai.com/en/articles/20001275
- ChatGPT desktop app: help.openai.com/en/articles/6825453
- Skills in ChatGPT and the Agent Skills standard: help.openai.com/en/articles/20001066
- Anthropic, import and export of memory: support.claude.com/en/articles/12123587
This is general practical guidance, not legal or compliance advice. Involve your own legal/compliance team and DPO. Status as of August 2026.