OpenAI has shipped a desktop app for Codex, its coding agent. Until now Codex lived in the terminal and on the web. As of this week it has a native macOS window: a list of agents, each in its own copy of the repository, a skills library, and a scheduler that runs jobs while you are not looking and files the results in an inbox.
We have run Claude Code and Cursor side by side on client builds for months. A third cockpit is not something we were looking for. But the fortnightly rhythm we promise every client has a specific shape, and the Codex app was built for that shape. So we put it on a live repository straight away. Here is what we found.
What Is in the Box
The app is a window over the Codex command-line agent, built with Electron so a Windows version can follow. OpenAI says that one is coming "very soon". The parts that matter:
- Parallel agents in worktrees. Start several agents at once. Each gets its own git worktree, so they cannot overwrite each other, and each lands as its own branch.
- Skills. Reusable instructions and tools the agent can pull in, the same idea Claude Code and Cursor have converged on.
- Automations. Jobs that run on a schedule in the background: triage new issues, look at why CI failed, draft a summary. Results go into a queue you review when you are ready. For now they run only while your Mac is awake; cloud-run automations are promised.
- GPT-5.2-Codex underneath. OpenAI's claim is that it tops TerminalBench for command-line work and lands roughly level with Gemini 3 and Claude Opus on SWE-bench, which is the bug-fixing benchmark closest to real work.
The commercial side is generous for now. Paid ChatGPT plans include it, free and Go accounts get it for a limited period, and paying users have their rate limits doubled while the promotion lasts. OpenAI says more than a million developers used Codex last month and usage has doubled since the December model release.
Simon Willison, who had early access, called it a solid app and made the sharper observation: it is a general agent harness disguised as a tool for programmers. Automations that read your inbox, fill a spreadsheet or triage tickets are the same feature as automations that fix your build.
Where It Fits in Takeoff & Climb
A build fortnight at Dev Force One starts with a set of specs from the roadmap and ends with a demo. In between, the work is parallel, and the overhead is coordination: which agent has which feature, what state each branch is in, what needs a human.
The Codex app makes that overhead visible. The agent list is the fortnight's board. One agent per spec, each on its own branch, each with a status. When one finishes, the diff is there to review and the branch is ready to merge. That is the same pattern we run in Cursor's agent view and in Claude Code with worktrees, so the question is not whether it works. It is whether this window is better at showing it.
For a single engineer juggling four branches, it is at least as good. The inbox for finished work is the piece the others do not have in the same form.
Where Automations Earn Their Keep
The feature we did not expect to like is the scheduler.
Cruise & Maintenance has a category of work that is important, small and easy to skip: dependency updates that need a look, a CI failure nobody has read, a changelog that should be drafted before the demo. These are exactly the jobs a scheduled agent should do and put in a queue for a human to approve.
We set three up on a client repository. A nightly job that checks the failed builds and writes a one-paragraph diagnosis. A weekly job that lists outdated packages with a recommendation on each. And a pre-demo job that drafts release notes from the merged branches. None of them merge anything. All of them save an hour of Monday.
The catch is the laptop. Automations run only while the machine is on, so a nightly job on a closed MacBook does not run. Until the cloud version arrives, this is a "while I am working" scheduler, not a true cron.
What We Are Cautious About
Three tools is two too many. Every agent harness wants its own rules file, its own skills format, its own permissions model. Keeping them consistent across a client repository is real work. Our approach is to keep one source of truth, the CLAUDE.md and spec files we already ship with every project, and treat any tool-specific configuration as generated from it.
Promotional pricing ends. Doubled rate limits and free-tier access are launch conditions, not a plan. We cost client builds on the standard tiers.
Permissions. A scheduler that can run shell commands on your machine while you are away needs the same scrutiny as any other agent. We keep automations read-only where possible and route anything that changes code through a pull request.
Mac only, today. Fine for our crew. Not fine for a client whose developers are on Windows, until that build ships.
The Bigger Picture
Three months ago the question was which coding agent to use. This week it is clearer that the agents are converging: parallel worktrees, skills, browser verification, scheduled jobs. The differences are in models and in the window you look at them through.
That is good news for founders. It means the process matters more than the vendor, and a good process moves between tools. The specs, the rules file and the review discipline are the assets. The cockpit is a preference.
Ready for a Build That Runs in Parallel?
At Dev Force One, every fortnight of Takeoff & Climb is a set of specs built in parallel and reviewed before you see the demo, whichever agent is flying that week.
What that gives you:
- Visible progress: one agent per roadmap item, one branch per feature, a demo every fortnight
- Maintenance that does not slip: scheduled checks on dependencies, builds and release notes in Cruise & Maintenance
- Process over vendor: your specs and rules travel with your code, not with a tool
Want to see it on your own idea? Book a Pre-flight call and let's plan the first fortnight.