OpenAI just dropped their Cursor killer

OpenAI’s new Codex app is winning developers by turning agent-based coding into a project-management workflow—one that keeps multiple workstreams moving at once, tracks progress in a UI, and reduces the “terminal tab paralysis” that slows real work. The core pitch is simple: Codex isn’t just another coding assistant. It’s a Mac-installable interface for orchestrating coding agents across repositories and branches, with shared history between the CLI and the app so work continues seamlessly across sessions and machines.

Codex starts as a CLI-backed system—models for coding plus a web layer for making pull requests—but the app adds the missing day-to-day control surface. Instead of micromanaging prompts in separate places, developers can run tasks as parallel “threads,” hop between them, and commit or open PRs when each thread finishes. The workflow is built around PR-ready outcomes: changes can be reviewed via diffs, committed with autogenerated messages, and pushed or turned into pull requests without leaving the Codex environment.

A major theme is speed and responsiveness. The transcript contrasts slow CI builds with a faster setup using Blacksmith (switching GitHub Actions runners from “Ubuntu latest” to “Blacksmith”), where job durations drop from minutes to under a minute for most runs. That same “less waiting, more visibility” mindset carries into Codex itself: long-running agent tasks can proceed while the developer works elsewhere, and the UI makes it clear what’s still in progress and what’s already completed.

Codex also tackles parallel development through a work-tree concept—Git-adjacent but implemented as actual copies that can later be synced back or turned into PRs. The creator is impressed by the thoughtfulness of the implementation but frustrated by its rough edges: work trees can start from the wrong branch, syncing back to local and pushing feels clunky, and environment-variable handling isn’t clearly surfaced. Still, the payoff is real: the same project can host multiple simultaneous tasks—one in a local work tree and another in a cloud environment—without agents stepping on the same code directory.

Beyond coding, Codex adds an ecosystem of “skills” and integrations. There’s an MCP-server angle, a skills browser with recommended tools (like deploying on Cloudflare, using Atlas for browser-based visibility, and updating Linear issues), and built-in image generation demonstrated through a one-prompt game-building flow. A standout “yeet” skill can stage, commit, and open a PR when the work is ready.

The transcript also draws a line between Codex and Cloud Code: Codex is positioned as the place for anything likely to be committed and pushed, while Cloud Code remains useful for one-off scripts, configuration tweaks, and general computer tasks not tightly bound to a git repo. Automations—prompt-driven cron-like jobs with file access—round out the system, spawning threads automatically to scan recent commits, summarize CI failures, or look for bugs on a schedule.

Finally, the creator claims Codex has displaced their usual IDE/terminal workflow so completely that they stopped using Cursor and even lost interest in opening full IDEs for line-by-line editing. The result is a shift from “commanding code” to orchestrating agents with a UI—so the developer’s job becomes supervising intent, reviewing outcomes, and shipping changes.