OpenAI plugs ChatGPT into Apple Messages on Mac

OpenAI has released an official plugin that connects ChatGPT to Apple Messages on the Mac. The practical upshot: the model can now read your conversation history, sort it, surface buried details, draft replies and send them for you — all from the chat interface. The plugin is also available to Codex and ChatGPT Work users, so it isn't limited to personal chats.

A short demo video the company circulated shows a user asking the chatbot to suggest follow-up messages to contacts based on the previous day's conversations. Beyond drafting, the plugin can delete messages, pull details from deep in the archive and edit texts before they go out. OpenAI frames these as productivity tools, but the line between "assistant that organizes" and "agent that acts for you" is thin when the model gets send permissions.

On privacy, OpenAI told Bloomberg the plugin runs locally on the device and does not build a full message index. The company did not say what exactly stays local, what gets sent to the cloud for processing, or what the retention policy looks like for conversation context shipped to the model. TechCrunch asked for clarification; the official answer remains a headline without a technical spec.

OpenAI itself warns against enabling persistent approval, which lets ChatGPT send messages without asking for confirmation each time. Turning that on "removes your last chance to review a message before it is sent in your name," the company says. The official guidance is to keep manual approval on and supervise every action the model takes — sensible advice that also undercuts the convenience the plugin is supposed to deliver.

The plugin is live now for users who link their Apple account, joining a wave of integrations that turn the chatbot into an active agent inside a walled garden. Searching and drafting inside message history is useful. But until OpenAI publishes a detailed architecture document — what is processed locally, what leaves the device, how long context is retained — users are being asked to trust broad assurances. Anyone who enables it today does so at their own risk, finger hovering over the approve button.