OpenAI has released a Codex desktop app for macOS, positioning it as a “command center” for running multiple coding agents in parallel and keeping long-running work organized by project.
The company says agents run in separate project threads so teams can switch tasks without losing context. The app also surfaces changes as diffs inside the thread so users can review and comment before moving work forward.
OpenAI’s developer documentation says the app is currently available on macOS for Apple Silicon. The company says users can sign in either with a ChatGPT account or an OpenAI API key, which matters for how enterprises separate subscription access from metered usage.
What OpenAI shipped
OpenAI’s product post emphasizes workflow structure over autocomplete. It says the Codex app organizes agents into separate threads by project, preserves context across tasks and lets users review the agent’s changes in-thread.
OpenAI also says users can open changes in an editor for manual edits, which keeps a human review loop at the center of the flow.
The app’s second pillar is isolation. OpenAI highlights built-in worktrees. It describes worktrees as a way to keep parallel agent work separated when multiple efforts touch the same repository.
In enterprise terms, this is meant as an attempt to reduce merge collisions and “agent sprawl” by defaulting to isolated workspaces rather than shared, mutable state.
Finally, OpenAI says the app supports scheduled “automations” that can run locally and deliver outputs into a review queue.
That pushes Codex from on-demand assistance toward recurring work like checks, updates or monitoring-style tasks that produce artifacts for approval.
Why the timing matters
Codex’s desktop move lands as vendors race to turn coding into the on-ramp for broader agent platforms. OpenAI itself describes Codex as a parallelized agent system for software work, with each task running in its own environment rather than as a single chat session.
The desktop client extends that approach into a daily workflow surface that can keep multiple agent threads active at once.
Competitive pressure is also easier to quantify than it was a year ago. Anthropic has said Claude Code reached $1 billion in run-rate revenue in six months, a milestone that signals real enterprise spend behind AI-assisted development workflows.
OpenAI’s Codex app launch reads as a response to that demand signal and to the risk of developer tooling budgets hardening around a small number of defaults.
Anthropic is also widening the frame beyond coding. Its Cowork research preview positions Claude as a broader “work” agent that can be granted access to a user-chosen local folder, then read, edit or create files in that folder.