If you read one line: OpenAI's Codex CLI v0.145.0 (21 July 2026) added an /import command that migrates your settings, MCP servers, plugins, sessions, commands, and project memories out of Cursor and Claude Code in one step. The news isn't the feature — it's what the feature is for: it deletes the switching cost that was the only real thing keeping teams on any one coding agent.

For two years the coding-agent market has been a leapfrog race on capability — a new model, a longer context window, a smarter sub-agent loop every few weeks. Founders learned to ignore most of it, because the thing that actually decided which agent you used wasn't the benchmark. It was the afternoon you'd already spent wiring up your MCP servers, rebuilding your custom commands, and letting the tool accumulate context about your codebase. That setup was the moat. Nobody re-tools every time a rival ships a point release.

Codex v0.145.0 goes straight at that moat. Per OpenAI's official release notes, the update "expanded /import to migrate Cursor and Claude Code settings, MCP servers, plugins, sessions, commands, and project-scoped memories" into Codex. In other words: the accumulated environment that made leaving painful now travels with you in a single command.

Why a migration command out-matters a new model#

Switching cost, not raw capability, is what keeps a team on a given agent. Models are fungible on a timescale of weeks; your configuration is not, because rebuilding it is manual, boring, and easy to keep postponing. A one-command import collapses that. The setup you'd have re-created by hand — the exact thing you'd cite as the reason not to bother trying a competitor — is now portable.

The moat around a coding agent was never the model. It was the config you'd already built. /import is a bulldozer aimed at that config.

Notice the direction, too. OpenAI is importing from Cursor and Claude Code — it's advertising that leaving those tools is easy. That's a competitor confident it wins the head-to-head once friction is removed, which tells you how the vendors themselves now rate lock-in as a defense: not much. And the logic cuts both ways. The same portability that pulls you into Codex is what lets you leave Codex later. Portability, not any one tool, is the feature.

The rest of v0.145.0, briefly#

The release did more than ship a migration path. It stabilized the opt-in multi-agent V2 experience — configurable sub-agent models, reasoning levels, and concurrency — so you can point cheaper models at cheaper sub-tasks. It added experimental Amazon Bedrock login with custom endpoints and GPT-5.6 Sol as the default Bedrock model, which matters for teams that must keep inference inside their own AWS account. And it added audio inputs, streaming realtime V3 conversations, and paginated thread history with search and resume. The current stable line is v0.145.0; the newest alpha, v0.146.0-alpha.6, is dated 24 July 2026 — so expect this to keep moving fast.

What a founder should actually do#

Not switch on reflex. The move is to re-run the evaluation you've been avoiding. If you've stayed on a coding agent partly because migrating felt like a project, that reason just expired. Decide on the axes that survive: which agent is genuinely better for your stack, which is cheaper at your token volume, which you trust with your code and your keys. We line those up in Claude Code vs Codex CLI vs Gemini CLI and, for the async side, Devin vs Codex vs Cursor vs Jules.

Then bank the general lesson, because it's bigger than one CLI. Lock-in built on setup friction is a decaying asset across the whole AI stack — the same principle behind keeping your LLM API choice portable and your agent memory exportable. When switching is one command, the winning posture isn't loyalty. It's staying ready to move — and letting the vendors compete for you every quarter.