OpenAI on May 21 released version 26.519 of its Codex desktop app for macOS, introducing Appshots—a hotkey-triggered feature that attaches the frontmost window to a conversation thread—and graduating goal mode from experimental to general availability across the desktop app, IDE extension, and command-line interface. The changes mark another step in the company's effort to turn its developer tool into a desktop-aware agent that can operate on tasks for long stretches.
The release, while incremental, extends the trajectory set by earlier Codex updates. In February, OpenAI positioned the app as a command center for agents; an April expansion added computer use, the ability to interact with desktop applications, an in-app browser, and automations. Appshots and goal mode now reduce two persistent friction points: gathering window-specific context and sustaining work toward a defined objective over time.
Appshots works by pressing both Command keys simultaneously—or a custom hotkey—capturing only the frontmost window, not the full desktop or background windows. It attaches a screenshot and any text the application makes available to the thread. For some widely used apps and web services, including Google Docs, Gmail, Google Sheets, and Google Slides, Codex may receive only the visible screenshot and not text that lies outside the scroll area, according to OpenAI's documentation. The feature requires that the user grant macOS Screen Recording and Accessibility permissions, and all captured content is stored locally within the session file, like any other attachment.
Goal mode, previously gated as an experiment, can now be started with the /goal command from any Codex surface. A user-provided goal text acts as both the initial prompt and the condition for completion; the mode provides controls to pause, resume, edit, and clear progress. OpenAI states that goal mode can drive toward a specific objective for "hours or days," though no independent benchmarks exist to verify reliability on tasks of that length.
Version 26.519 also delivers advanced in-app browser annotations and browser reliability improvements. Locked computer use—Codex's ability to continue operating allowed desktop applications after the Mac locks—is part of the release, but it comes with conditions: the user must explicitly opt in, the feature is limited to eligible Computer Use users, and it operates under safeguards and regional restrictions. For ChatGPT Business customers, plugin sharing through marketplace sources is now available; Enterprise support is planned but not yet live.
Taken together, the updates fill in a pattern. OpenAI introduced the Codex desktop app in February, extended it to Windows in March, and added computer use, desktop-app interaction, and automations in April. Each release has widened the tool's access to the local machine environment and lengthened the horizon over which it can work. May's version makes that machine awareness more granular—one window at a time—and formally commits to goal mode as a core capability.
The new features arrive with significant qualifiers. Appshots captures a single frontmost window and cannot guarantee extraction of all off-screen text; for several Google productivity apps, only the visible portion is available. Goal mode's long-duration capability remains an OpenAI claim, not an independently verified benchmark. The release notes do not specify which plans (Plus, Pro, Team, Enterprise, Edu) or regions (including the EU and UK) will have access to Appshots, goal mode, or locked computer use at launch. Privacy and data-retention terms for window content—whether text from an Appshot is used for model training by default—are not detailed in the available documentation. No dedicated blog post or executive statement accompanied the release; the features were described in the developer changelog and help centre notes. The sources also do not address if Appshots or goal mode will appear on Windows.
For developers building on Codex, the update reduces the manual work of pasting context across tools and offers a more persistent mode for tackling tasks that span hours. It sharpens the competitive dynamic among AI coding assistants, which are increasingly measured by how deeply they integrate into a user's local workflow.