OpenAI Publishes Codex Desktop App Guide
07 Jul 2026
OpenAI has published a getting-started guide for its Codex desktop app, laying out how the tool organizes coding work into local, folder-based projects rather than granting open-ended access to a user's computer.
What the guide covers
According to the report, Codex is a desktop application that users download and sign into with a ChatGPT account. Once inside, the core interface includes a sidebar menu, projects, settings, and a chat window, along with a search function for locating previous tasks and projects.
The basic unit of work is a thread — described as similar to a ChatGPT chat, a space for back-and-forth interaction to accomplish a task. Threads can stand alone or be nested inside a project, which is tied to a specific folder on the user's computer. Notably, project folders don't need to contain any files to start: users can drag existing files in for Codex to work with, or leave the folder empty and let Codex generate new files from scratch.
Codex can inspect files, create new ones, edit documents, organize information, and build things within that scoped environment. When a project is set to "Work locally," Codex is limited to operating within its designated folder using selected tools — the guide states plainly that Codex does not automatically get access to everything on a user's computer.
Settings and controls
The app includes Default and Full permissions settings, with Full permissions giving Codex more latitude to act on advanced tasks. Other settings govern personalization and whether Codex continues working while the computer sleeps. Users can also select a reasoning level, with a default recommended model and the option to increase reasoning for harder tasks — though the report gives no specifics on which models or reasoning tiers are available.
While a task is running, users can invoke a Steer function to course-correct Codex in real time. Multiple Codex tasks can also run simultaneously without interrupting one another. Plugins are available as well, intended to support specific work or repeatable processes, though the guide does not detail what plugins exist or how they're built.
Risks and gaps flagged in the report
A few caveats stand out. Codex may stop running if the computer goes to sleep, potentially interrupting in-progress tasks. Full permissions settings increase Codex's room to act, which could raise exposure if tasks go unmonitored. And the guide offers limited detail on the technical boundaries of "Work locally" mode, leaving some ambiguity about how contained the sandbox truly is.
The report also notes several gaps: there's no information on which operating systems support the desktop app, no pricing or subscription details, no specifics on reasoning levels or model names, and no explanation of how permissions are technically enforced. No quantitative metrics — pricing, user counts, or dates — appear anywhere in the sources.
Why founders should care
For early-stage teams evaluating AI coding tools, this structure likely signals that Codex is built for scoped, task-specific automation rather than broad system access — a design choice that may reduce risk for teams cautious about giving AI tools free rein. The Default/Full permissions split suggests founders could calibrate autonomy based on how sensitive a given task is, though without technical detail on enforcement, teams should probably treat permission settings as a starting point for caution rather than a guarantee.
The ability to run multiple threads at once could plausibly support parallelized engineering or research workflows, which may appeal to lean teams trying to move faster without adding headcount. At the same time, the sleep-related interruption risk suggests founders relying on long-running Codex tasks may need to think about always-on infrastructure to avoid dropped work.
Finally, plugin support hints at an emerging ecosystem for customizing Codex to startup-specific workflows — though with no details yet on what plugins exist, it's too early to gauge how mature or useful that ecosystem will be.
Bottom line
OpenAI's guide clarifies how Codex is meant to be used — locally scoped, permission-gated, and thread-based — but leaves open real questions about platform support, pricing, and the technical rigor of its security boundaries. Founders experimenting with Codex should weigh its parallel-task and steering capabilities against the current lack of clarity on containment and infrastructure requirements.