# Codex App
The Codex App is the desktop surface of [[OpenAI Codex]]; a full-featured GUI built around the same agent runtime as [[Codex CLI]] and Codex Cloud, but oriented toward longer sessions and richer interaction patterns. It is what someone running multiple parallel agent threads on a project would reach for instead of the CLI.
It is the closest thing OpenAI ships to a full agentic IDE; not a code editor itself, but a control plane for AI-driven development that integrates with the rest of the toolchain.
## Layout
Three primary panes:
- **Project sidebar**; pinned repos and environments, navigation across active work.
- **Thread list**; concurrent agent conversations on the current project, with status and progress.
- **Review pane**; the diff/preview surface where changes are inspected before being applied or committed.
This split is the App's central abstraction; threads on the left, the work product on the right, projects above. It maps closely to how a human reviews multiple parallel pull requests.
## Capabilities
- **Local environments**; run agent jobs against the local filesystem, like the CLI, but with the GUI affordances (drag-and-drop, image attach, visual diff).
- **In-app browser**; an embedded browser the agent can drive for research, scraping, and verification tasks.
- **Computer use**; agent control over GUI applications, mirroring [[Claude Computer use]].
- **Automations**; saved workflows triggered on schedules or events.
- **Worktrees**; first-class git worktree support so multiple agent threads can operate on the same repo on different branches without stepping on each other.
- **Cloud handoff**; jobs can be promoted to Codex Cloud for async / heavier execution.
## When to choose the App over the CLI
- Multiple parallel agents on the same project; the thread list and worktree integration handle this far better than a tiled terminal.
- Heavy use of computer use or in-app browser; CLI surfaces cannot match.
- Visual review of large diffs; the review pane is purpose-built for this.
- Non-terminal collaborators; the App is approachable for users not living in tmux.
The CLI remains better for scripting, CI integration, fast iteration on a single thread, and unattended automation.
## Authentication and config
Shared with the CLI and Cloud surfaces. The App reads `.codex/config.toml` and the same auth tokens, so a session started in the CLI can be resumed in the App and vice versa.
## Practical patterns
- **One thread per concern**; treat threads like browser tabs. Open a thread for the failing test, another for the refactor, another for the doc update. Reviewing them in the diff pane in turn beats juggling them in a single conversation.
- **Worktrees as concurrency**; spin a worktree per active thread instead of letting agents fight over the working tree. The App makes this cheap; the CLI does not.
- **Computer use sparingly**; for any task that has a programmatic alternative (curl, gh, a script), use the alternative. Computer use is the right tool for genuinely GUI-only flows; everything else is faster and more reliable through actual APIs.
- **Promote, don't migrate**; long-running threads belong in [[Codex Cloud]]. Keep the App for interactive work that benefits from human feedback per turn.
- **Automations as ratchets**; once a workflow has been done by hand twice, save it as an automation. The App's automation surface is the lowest-friction path to "this should run on a schedule" without touching cron or n8n.
## Trade-offs vs the CLI
- The App carries a desktop-app footprint; memory and disk cost are real on smaller machines.
- The App is harder to script; there is no `codex-app exec --json` equivalent. Pipelines belong in the CLI.
- Plugin/feature parity lags the CLI by a release or two on average.
## References
- Codex product family: https://openai.com/codex/
- Documentation: https://developers.openai.com/codex/
## Related
- [[OpenAI Codex]]
- [[Codex CLI]]
- [[Codex plugin for Claude Code]]
- [[Claude Code]]
- [[AI Agent Harness]]
- [[AI Agents]]