AI library

Source · docs

Codex docs: Import from another agent

Author
OpenAI
Added
5 November 2026
Original
https://learn.chatgpt.com/docs/import

Summary

OpenAI's documentation page for importing an existing setup from another coding agent into Codex (formerly at developers.openai.com/codex/import). The ChatGPT desktop app imports from Claude Code, Claude Cowork, or Cursor; Codex CLI imports from Claude Code or Cursor. Instructions, settings, skills, plugins, MCP servers, hooks, subagents, project memories, and recent chats are each mapped to the Codex equivalent. The page shows no publication date; it was read on 2026-11-05.

Key points

  • How to run it: in Codex CLI, type /import in an idle session, pick Claude Code or Cursor, choose the setup, project files, and chats, then review the result. In the desktop app, use Settings > Import.
  • Mapping: instruction files → AGENTS.md; settings.json → config.toml; skills → skills; slash commands → skills; plugins → plugins; MCP server configuration → Codex MCP configuration; hooks → Codex hooks; subagents → Codex subagents; "Project memories from Claude Code" → Codex memories; chats from the last 30 days → chats.
  • One-time vs sync: a one-time copy by default. Only the desktop app offers automatic updates (under Settings > Import) to keep imported work in sync; the page does not describe sync for the CLI.
  • Limits: the CLI imports at most 50 chats from the last 30 days. Importing does not change or delete the source agent's setup.
  • Not stated: which exact files count as Claude Code "project memories" (for example CLAUDE.md versus the auto-memory MEMORY.md directory), and what happens to an existing AGENTS.md. A secondary write-up says an existing AGENTS.md is skipped; that is unconfirmed.
  • Check after importing: tool permissions in imported skills and agents, MCP servers with custom auth, headers, or environment variables (may need signing in again), hooks whose behaviour may differ, and prompt templates that rely on arguments or shell interpolation.
  • Codex stores its own memories under ~/.codex ($CODEX_HOME/memories_v2/ in a fresh CLI home), per the separate Memories page.

Why it matters

It is the most complete cross-agent migration path so far: almost every Claude Code customisation (skills, hooks, MCP, subagents, memories) has a Codex counterpart, which makes trying a second agent on the same project cheap.