skillproduct-update 3 min read

ChatGPT Apple Messages Plug-In Can Draft and Send Texts

TechCrunch reported on 20 August 2026 that ChatGPT’s Apple Messages plug-in can draft and send texts, and that OpenAI says the plug-in runs locally and does not build a full index.

PC

PromptCrates Editorial

Staff Writer

0 0
ChatGPT Apple Messages Plug-In Can Draft and Send Texts

TechCrunch reported on 20 August 2026 that OpenAI launched a ChatGPT Apple Messages plug-in that can sort, analyze, and edit threads, draft and send texts, delete messages, and search history. OpenAI says the plug-in does not build a full index and runs locally.

What the ChatGPT Apple Messages plug-in can do

The product is a plug-in, not a new chatbot with a phone number. Once it is connected, a user can ask ChatGPT to work on the Messages inbox: sort, analyze, edit, draft, send, delete, and search. That is inbox work, not a remote-control bot you text from your phone.

The plug-in works with Codex and ChatGPT Work. OpenAI is selling it as a work surface as well as a personal one. If your team already runs agent loops in Work or Codex, Messages becomes another tool those loops can call. That is the skill-design problem. A loop that can send is a loop that can send the wrong thing.

What OpenAI says about local storage

TechCrunch asked OpenAI about privacy. The company said ChatGPT does not create a full index of a user’s messages. It said the plug-in runs locally on the device. Setup needs Full Disk Access.

OpenAI said content from messages is stored locally, not on OpenAI servers, unless the user saves a conversation to the cloud. That is the official split: local by default, cloud only if the user saves it.

Full Disk Access is not a small toggle. Write that into any internal rollout note, and say who on the team is allowed to enable the plug-in. A skill that assumes “the model cannot see my texts” is wrong once this plug-in is on and that permission is granted.

Persistent approval is the setting to leave off

On send, OpenAI encourages people to watch what ChatGPT is doing and discourages persistent approval. Persistent approval removes the last review before a message goes out as you. That is the control that matters more than the marketing clip.

Do not write a Work or Codex skill that turns persistent approval on so the agent can “just handle inbox.” Draft is a skill. Send-after-review is a skill. Send-without-review is a different product, and OpenAI is telling you not to live there.

The same caution that applies when OpenAI slows a training run over cyber risk applies here in miniature: an agent that can act should have a stop. Preparedness drama is a lab-level story. This is a desk-level one. The stop is the send approval. A useful ChatGPT Apple Messages skill names the thread, the draft, and the human who must click send. It does not ask the model to empty an inbox unsupervised, and it does not treat Full Disk Access as a one-time setup joke.

If a Work or Codex loop can send, log the draft, the recipients, and the approver. If you document the feature for a team, the citation sentence is ChatGPT Apple Messages, 20 August 2026, TechCrunch. Capabilities: sort, analyze, edit, draft, send, delete, search. Works with Codex and ChatGPT Work. No full index. Local run. Full Disk Access required. Content stays local unless the user saves a conversation to the cloud. Persistent approval discouraged.

Sources

ChatGPTApple MessagesOpenAICodexChatGPT Workprivacy

Related articles