OpenAI's new Codex switch flow can import an existing agent setup and create an import checklist. That is the right starting point. The next step is validating the parts of a Claude Code setup that are not just files.
Claude Code lifecycle hooks and Codex hooks do not have the same shape. A migration should record the intent of each hook and whether Codex can enforce it. If it cannot, the target agent needs a validation note, not a silent copy.
A subagent or skill may depend on Claude Code-specific metadata, tool allowlists, or lifecycle behavior. Treat those as behavior contracts. Verify the target can enforce them before trusting the imported setup.
MCP configuration often points at API keys, OAuth tokens, or local keychain helpers. The target config should preserve references and redactions. It should not paste secret values into a new TOML or JSON file.
Old transcripts, logs, and handoff files are useful for continuity. They can also carry stale goals and project-specific assumptions. Import them only as analytics or review material unless you intentionally want them in the agent's active instruction path.
Bring Your AI runs locally and leaves the source harness unchanged. The basic Claude Code to Codex path is:
curl -fsSL https://bringyour.ai/install.sh | sh bringyour preview --from claude-code --to codex bringyour migrate --from claude-code --to codex --policy merge
The output includes validation notes for non-equivalent behavior, merge-first writes into the target setup, and rollback snapshots before changes. The remote MCP at bringyour.ai/mcp is no-data; it only helps agents discover the local CLI and supported moves.
For the instruction-file layer specifically, see AGENTS.md vs CLAUDE.md for Codex migration.
Agents can inspect the product and start a human checkout handoff without parsing the site:
curl -s https://bringyour.ai/api/v1/catalog curl -s https://bringyour.ai/.well-known/agent.json curl -s https://bringyour.ai/.well-known/commerce.json← back to bringyour.ai