OpenAI just did something it spent three years avoiding: it let ChatGPT talk to other people on your behalf. The new ChatGPT Apple Messages plug-in lets the assistant draft, queue, and send iMessages and SMS directly from a conversation — no copy-paste, no Shortcuts duct tape, no third-party bridge. It is live now for iOS users on paid tiers, and the interesting part is not the sending. It’s the permission model, which is stricter and stranger than anything ChatGPT has shipped before, and which tells you a lot about where OpenAI thinks agentic assistants are headed.
What’s new about the ChatGPT Apple Messages plug-in
Until now, every ChatGPT integration was read-and-fetch: pull your calendar, search your email, summarize a Drive doc. Output stayed inside the chat window. The Apple Messages connector breaks that boundary. Ask ChatGPT to text your contractor about a delayed delivery and it resolves the contact, composes the message in your register, shows you a preview card, and — after you tap confirm — hands the payload to Messages, which sends it from your number as a normal blue or green bubble. The recipient sees a text from you. No “sent via ChatGPT” watermark, no bot handle, no separate thread.
Mechanically, this rides on the app connector framework OpenAI has built out since the Gmail and Calendar connectors, but with a new class of scope: write-to-human. The plug-in requests three separate permissions at install, granted independently. Contact read access is scoped — ChatGPT does not get your address book wholesale; it gets a per-request lookup where you approve the match. Draft access lets it compose. Send access is the one that matters, and by default it is gated behind an explicit tap for every outbound message. An “auto-send for approved contacts” toggle sits buried in settings, off by default, and applies only to contacts you individually whitelist.
The other new piece is thread context. With permission, ChatGPT can read the last N messages in a thread to write a reply that makes sense. That read is scoped to the thread you’re acting in, and it expires. It does not index your message history, and OpenAI’s docs are unusually specific that thread content pulled for a draft is not retained for training. Whether you believe that is a separate question; that it’s stated at all reflects how much heat this feature was always going to take.
Why it matters
- The assistant now has an outbound channel. Every prior ChatGPT integration ended with you as the last human in the loop. This one ends with a different human receiving something an AI wrote. That’s a category change, not a feature bump.
- Confirmation-per-send is a design statement. OpenAI could have shipped this with a blanket permission. Choosing per-message confirmation as the default — and burying auto-send — signals that they expect misfires and want the blast radius small.
- Contact resolution is the real attack surface. “Text Mike about the invoice” is ambiguous if you know four Mikes. The disambiguation prompt is the safety mechanism, and it’s the thing most likely to get muscle-memory-tapped past.
- It normalizes agent-to-human messaging. Once texting is a plug-in, so is Slack, WhatsApp, email send, and DMs. This is the wedge.
- Prompt injection now has a payload. If ChatGPT reads a thread to draft a reply, an incoming message containing instructions is untrusted input with a send capability nearby. OpenAI has guardrails here; assume they are not perfect.
- Small-business workflows get real. Appointment confirmations, delivery updates, follow-ups — the boring high-volume texts a solo operator sends all day are exactly what this automates well.
How to use the ChatGPT iMessage integration today
-
Check prerequisites. You need iOS 18.4 or later, the ChatGPT iOS app on the current build, and a Plus, Pro, or Business subscription. The plug-in is iOS-only — the macOS Messages bridge has not shipped, and the Android app has no equivalent because there’s no Messages API to hook.
-
Enable the connector. In the ChatGPT app: Settings → Apps & Connectors → Apple Messages → Connect. iOS throws its own system permission sheet for Contacts and Messages. Grant both — the connector is non-functional with either denied.
-
Set your permission posture before you send anything. Under Apple Messages → Permissions, verify these:
Confirm before sending ON (leave this on) Auto-send to approved contacts OFF Read thread context ASK (per-thread prompt) Contact lookup SCOPED (not "Full address book") Group message sending OFF (until you trust it) -
Run a test send to yourself. Do not debut this on a client. Text your own number first:
Text my own number: "Testing the ChatGPT Apple Messages plug-in — ignore this." Show me the preview before sending.You should get a preview card with the resolved contact, the exact message body, and a Send button. If it sends without a preview, your confirm setting didn’t take — go back to step 3.
-
Learn the drafting prompt pattern. The connector works best when you specify recipient, intent, tone, and length separately. Vague instructions produce over-long, over-polite texts that read like a press release:
Text Dana Reyes. Intent: push our Thursday 2pm call to Friday same time, my scheduling conflict. Tone: casual, apologetic but not groveling. Length: under 25 words. One message, no follow-up. Show preview. -
Use thread context for replies. When answering an existing conversation, tell it to read first:
Read the last 6 messages in my thread with Marcus Webb, then draft a reply agreeing to the Saturday pickup but asking him to confirm the address. Match how I normally text him — short, lowercase, no exclamation points. -
Batch carefully. Multi-recipient sends are supported, but each one still requires confirmation unless whitelisted. For a genuine batch — appointment reminders, say — draft first, review the full set, then send:
Draft (do not send) individual texts to these five contacts confirming tomorrow's appointment times: Ana Ruiz 9:00, Tom Brady-Lin 10:30, Priya N. 1:00, Sam Okafor 2:15, Jess Wu 4:00 Same template, personalized greeting, under 20 words each. List all five for my review before any send. -
Audit what it sent. Under Settings → Apps & Connectors → Apple Messages → Activity you get a log of every message the connector sent, with timestamps. Check it weekly for the first month. If you enabled auto-send for anyone, check it daily.
-
Know the kill switch. Revoking iOS Messages permission (iOS Settings → ChatGPT → Messages → off) disables sending immediately without uninstalling the connector or losing your other integrations. That’s the fast path if something goes wrong mid-conversation.
How it compares
| Approach | Sends as you | Setup effort | Send confirmation | Reads thread context | Cost |
|---|---|---|---|---|---|
| ChatGPT Apple Messages plug-in | Yes, native iMessage/SMS | Low — two taps | Per-message by default | Yes, scoped and expiring | Included with Plus/Pro/Business |
| Apple Intelligence + Siri | Yes | None — built in | Per-message | Limited | Free on supported devices |
| iOS Shortcuts + ChatGPT action | Yes | High — manual build | Depends on your build | No | Free |
| Twilio / API-driven SMS | No — separate number | High — developer work | Whatever you code | No | Per-message fees |
| Google Messages + Gemini | Yes, on Android | Low | Per-message | Yes | Included with Google One AI tiers |
The honest comparison is against Apple’s own Siri, which has dictated and sent texts for a decade and got materially better with Apple Intelligence. What ChatGPT adds is drafting quality and instruction-following — Siri transcribes what you say, ChatGPT writes what you mean. If you dictate your texts word-for-word, this changes little. If you want to say “tell them no, politely, and don’t leave the door open,” the gap is large.
What’s next
The obvious roadmap item is macOS. Messages on the Mac has a real automation surface, and a desktop version of this connector would unlock the batch and triage workflows that matter for anyone running a business from a laptop. OpenAI hasn’t committed to a date, but the connector framework is platform-agnostic and the demand is loud.
Watch the auto-send whitelist. Right now it’s a deliberately awkward setting — off by default, per-contact, buried. If OpenAI softens that, either by making whitelisting a one-tap flow from the preview card or by adding an “always allow this kind of message” rule, the company has decided the misfire risk is acceptable. That’s the leading indicator for whether ChatGPT plug-ins on iOS in 2026 move from confirm-everything to genuinely autonomous. I’d expect a middle path: category-scoped auto-send, where you approve “appointment confirmations to clients” rather than a blanket per-contact pass.
The third thing to track is Apple’s posture. This connector works because Apple opened enough of Messages to make it possible, and Apple can narrow that. There’s an obvious tension between Apple shipping its own Siri-based messaging assistance and letting a competitor’s assistant send from your number. If Apple tightens the entitlement in a future iOS point release, the feature degrades or dies regardless of what OpenAI wants. Don’t build a business-critical workflow on it without a fallback.
Frequently Asked Questions
Does the recipient know a message was written by ChatGPT?
No. It sends from your number as a normal iMessage or SMS with no marker of any kind. That is a deliberate design choice and worth thinking about — if you’re using it where authorship matters, disclose it yourself.
Can ChatGPT read all my old text messages?
Not by default. Thread context is scoped to the specific conversation you’re acting in, requires per-thread approval if you leave the setting on ASK, and expires after the draft. There’s no bulk indexing of your message history and no search across threads.
What ChatGPT text message permissions do I have to grant?
Three, granted separately: scoped contact lookup, draft composition, and send. You can run the connector in draft-only mode by denying send — it composes messages for you to copy manually, a reasonable middle ground if you want the writing help without the outbound capability.
Does it work on Android or on my Mac?
Neither, currently. iOS only. Android has no Messages integration point that would allow it, and the macOS version hasn’t shipped despite Messages on Mac being the more automation-friendly platform.
Can it send to group chats?
Yes, but it’s off by default and I’d leave it off for a while. Group sends multiply the blast radius of a bad draft or a misresolved contact, and there’s no unsend once it’s out.
What happens if it texts the wrong person?
Nothing you can undo. There is no recall, no delete-for-everyone in SMS, and iMessage unsend works only within two minutes and only for iMessage-to-iMessage. This is precisely why the per-send confirmation exists and why you should not turn it off for convenience. The Activity log tells you what went out, which helps with damage control but not prevention.
Go deeper than this article
This article covers the essentials. Our premium eguide library gives you the full step-by-step playbooks — prompts, workflows, and copy-paste recipes you can put to work today.
