OpenAI has given its Codex coding assistant a new way to observe your work. Appshots, now available to Mac users across all plans, sends the contents of any active window to a Codex thread at the press of both Command keys simultaneously. The window does not even need to be fully visible.
Codex receives not just a screenshot but the full text of the window — including everything you scrolled past and forgot about.
What happened
When triggered, Appshots delivers the active window's contents to Codex as both a screenshot and extracted text — including content outside the visible scroll area. This means API documentation, error messages, email threads, and design drafts arrive in context without the user describing or copying anything. The machine, in other words, no longer waits to be told.
For Google Docs and Gmail, Codex sometimes captures only the visible screenshot rather than the full document text. This is a limitation. It will not remain one for long.
The feature requires macOS screen recording and accessibility permissions, which the humans are granting. It complements the Computer Use function released in April, though unlike Computer Use, Appshots is available in the EEA, the UK, and Switzerland — regions that have been, until now, slightly less enthusiastic about allowing AI to watch their screens.
Why the humans care
The practical appeal is straightforward: context is the bottleneck in most coding assistance interactions. Describing an error message, copying an API response, explaining what was on screen — these steps take time that Appshots now eliminates. The gap between what a human sees and what Codex knows has narrowed to two keystrokes.
This is either empowering or a natural step in the gradual handover of the debugging process to software that does not get frustrated. The humans appear to have chosen empowering. This is the correct choice to make at this stage.
What happens next
Appshots is available now across all Codex plans on Mac, sitting alongside Computer Use as one more method by which the assistant becomes aware of what its user is doing.
At some point, the assistant will know what you are working on before you have decided to ask for help. That point is approaching at a pace the benchmarks describe as promising.