A full conversation,
on the model you pick.
The three buttons above the keyboard work inside one thread, on one reply. For work bigger than that, the app has a dedicated agent chat — a standing conversation that runs on whichever model you choose. Apple Intelligence handles it on your iPhone, with no key and no credit to set up; Private Cloud Compute takes a turn that needs a bigger context window; Claude takes the long multi-step work, on prepaid AI Credit or your own Anthropic key. Ask it anything, with your own documents, the web, and your message history already loaded, and send the answer straight back into the chat it came from.
Everything is set per conversation, right in the chat: the model — and, on Claude, how hard it thinks — which tools the agent may use, the active Skill, and, most importantly, the scope. You pick which of your conversations the agent may read, and everything outside that selection is invisible to it. Locked chats never appear at all.
Know your way around
Everything in the agent chat is one tap away. Tap a feature to see where it lives and what it opens.
What it adds over an in-thread draft
Most of it is the chat itself, and holds on any engine you pick.
The rest is the Claude path. There is no on-device counterpart for any of it, so while Apple Intelligence is selected the app removes those controls from the chat rather than greying them out — the sheets are simply shorter.
Extended Thinking, visibleSub-agents in parallelPowerPoint, Excel, PDF & Word creation — own keyCited sourcesMemory and attachmentsUp to 1M tokens of context on Sonnet 5.5 and Opus 5.5The model is yours to pick, in the chat itself, and the chat remembers which one you used last — with no Claude connection set up it opens on Apple Intelligence. Once a conversation has started on Claude it stays there; start a new topic to go back to Apple Intelligence. What each one costs, and how they compare →
It proposes. You approve.
Wherever the AI runs — the Skill button or the agent chat — it can't touch your conversations or your device directly. When it wants to act, it must produce an action card — a structured preview of exactly what would be sent or created. A drafted message lands in the composer of the chat it is addressed to, so the send stays yours; the other messaging cards open a recipient picker; device cards file into iOS with a single tap; Agent Skills cards carry the documents Claude creates — and nothing moves until you say so. Nobody on the other end ever receives something you have not read.
Sarah ChenEvery card it can put in front of you
Four groups, by what a card does when you tap it. Every one is a full preview before anything happens, every one shows its own status as it works, and every one stays in the thread afterwards, so you can look back at what was sent. Tap a card to see what it holds.
Something to send. A draft opens the composer of the chat it is for; the rest open the recipient picker.
Something to file away. One tap writes it into the iOS app it belongs in.
A document Claude built in its sandbox, to preview, share or save.
The one card that is not the end of something — it asks, and the answer carries on the conversation.
Draft Message
What it is
The plainest card of the eighteen: a message Claude has written, addressed to somebody, waiting for you to read it and say where it goes. It is the only card in the messaging group that is purely text.
What the card shows you
- Recipient
- The chat Claude addressed it to — changeable, and not final until you confirm in the picker.
- Platform
- WhatsApp, Telegram, Instagram or Messenger.
- Message
- The text itself, with its formatting intact — bold, italic, links and mentions all survive the trip.
- Forum topic
- On a Telegram supergroup, the topic the message will be posted in.
When you tap it
Opens the recipient picker. Nothing is sent until you confirm there.
You choose who, after you have read it
Tapping the card opens a multi-select recipient picker — send to one chat, or fan out to many at once. The card then shows its real-time status: idle, sending, sent with a recipient count, or failed. A partial send reads “Sent to 2 of 3 chats” and names the first error, and Retry pre-selects only the recipients that failed, so nobody is messaged twice.
See also
Image Message
What it is
A photo on its way into a chat, with the caption Claude wrote for it. The picture is one you supplied — Claude describes it and captions it, it does not draw it.
What the card shows you
- Image
- The file, with its name, size, type and pixel dimensions.
- Caption
- The text that travels with the photo.
- Description
- What Claude understood the picture to be — which is how it knew the caption was right.
- Recipient
- The chat and platform it is addressed to.
When you tap it
Opens the recipient picker, then sends the photo and its caption together.
You choose who, after you have read it
Tapping the card opens a multi-select recipient picker — send to one chat, or fan out to many at once. The card then shows its real-time status: idle, sending, sent with a recipient count, or failed. A partial send reads “Sent to 2 of 3 chats” and names the first error, and Retry pre-selects only the recipients that failed, so nobody is messaged twice.
The media is yours, not invented
This card references an attachment you provided earlier in the conversation. Claude describes and captions it, and can decide where it should go — but it never fabricates the file itself, so what arrives is the bytes you supplied.
See also
Video Message
What it is
A video on its way into a chat, with its caption.
What the card shows you
- Video
- The file, with its name, size, type, dimensions and duration.
- Caption
- The text that travels with the video.
- Recipient
- The chat and platform it is addressed to.
When you tap it
Opens the recipient picker, then sends the video and its caption together.
You choose who, after you have read it
Tapping the card opens a multi-select recipient picker — send to one chat, or fan out to many at once. The card then shows its real-time status: idle, sending, sent with a recipient count, or failed. A partial send reads “Sent to 2 of 3 chats” and names the first error, and Retry pre-selects only the recipients that failed, so nobody is messaged twice.
The media is yours, not invented
This card references an attachment you provided earlier in the conversation. Claude describes and captions it, and can decide where it should go — but it never fabricates the file itself, so what arrives is the bytes you supplied.
See also
Audio Message
What it is
An audio file on its way into a chat. It is either an ordinary audio attachment or a voice note, and the card says which — the two arrive looking quite different at the other end.
What the card shows you
- Audio
- The file, with its name, size, type and duration.
- Voice note
- Whether it will arrive as a recorded voice message rather than as a plain audio file.
- Recipient
- The chat and platform it is addressed to.
When you tap it
Opens the recipient picker, then sends the audio.
You choose who, after you have read it
Tapping the card opens a multi-select recipient picker — send to one chat, or fan out to many at once. The card then shows its real-time status: idle, sending, sent with a recipient count, or failed. A partial send reads “Sent to 2 of 3 chats” and names the first error, and Retry pre-selects only the recipients that failed, so nobody is messaged twice.
The media is yours, not invented
This card references an attachment you provided earlier in the conversation. Claude describes and captions it, and can decide where it should go — but it never fabricates the file itself, so what arrives is the bytes you supplied.
See also
File Message
What it is
A document on its way into a chat — a PDF, a spreadsheet, a deck, anything you attached — with a caption.
What the card shows you
- File
- The document, with its name, size and type.
- Caption
- The text that travels with it.
- Recipient
- The chat and platform it is addressed to.
When you tap it
Opens the recipient picker, then sends the document.
You choose who, after you have read it
Tapping the card opens a multi-select recipient picker — send to one chat, or fan out to many at once. The card then shows its real-time status: idle, sending, sent with a recipient count, or failed. A partial send reads “Sent to 2 of 3 chats” and names the first error, and Retry pre-selects only the recipients that failed, so nobody is messaged twice.
The media is yours, not invented
This card references an attachment you provided earlier in the conversation. Claude describes and captions it, and can decide where it should go — but it never fabricates the file itself, so what arrives is the bytes you supplied.
See also
Watch out for
Instagram takes images, video and audio, but not documents. A file card addressed to an Instagram chat has nowhere to land — send it on one of the other three networks.
Contact Message
What it is
Somebody’s details, shared into a chat as a contact card rather than typed out as text. Claude fills in the vCard from what it found, and you check it before it goes.
What the card shows you
- Name
- First and last name, and the display name the card carries.
- Reaching them
- Phone number, email address and username, where each is known.
- Work
- Company and job title.
- Business details
- A business website, email and address, for a contact that has them.
- Recipient
- The chat and platform the card is addressed to.
When you tap it
Opens the recipient picker, then shares the contact card.
You choose who, after you have read it
Tapping the card opens a multi-select recipient picker — send to one chat, or fan out to many at once. The card then shows its real-time status: idle, sending, sent with a recipient count, or failed. A partial send reads “Sent to 2 of 3 chats” and names the first error, and Retry pre-selects only the recipients that failed, so nobody is messaged twice.
See also
Event Message
What it is
A calendar event shared into a chat, so the other person can add it to their own calendar. This is the sharing card — the one that files an event into your calendar is Event, in the device group.
What the card shows you
- Title
- What the event is called.
- When
- Its start and end.
- Where
- The location, when there is one.
- Notes
- Anything else that belongs with it.
- Recipient
- The chat and platform it is addressed to.
When you tap it
Opens the recipient picker, then shares the event.
You choose who, after you have read it
Tapping the card opens a multi-select recipient picker — send to one chat, or fan out to many at once. The card then shows its real-time status: idle, sending, sent with a recipient count, or failed. A partial send reads “Sent to 2 of 3 chats” and names the first error, and Retry pre-selects only the recipients that failed, so nobody is messaged twice.
It travels two ways
Where a network can render an event natively, it arrives as one. Where it cannot, the same event is sent as an .ics calendar file, which every calendar app opens — so the event survives either way.
See also
Location Message
What it is
A place, shared into a chat as a map pin rather than as an address somebody has to copy out.
What the card shows you
- Place
- The name of the place.
- Address
- Its street address.
- Coordinates
- The latitude and longitude the pin drops on.
- Recipient
- The chat and platform it is addressed to.
When you tap it
Opens the recipient picker, then shares the pin.
You choose who, after you have read it
Tapping the card opens a multi-select recipient picker — send to one chat, or fan out to many at once. The card then shows its real-time status: idle, sending, sent with a recipient count, or failed. A partial send reads “Sent to 2 of 3 chats” and names the first error, and Retry pre-selects only the recipients that failed, so nobody is messaged twice.
See also
Poll Message
What it is
A poll, composed and ready to run — the question, the options, and the rules for answering it. The richest card of the eighteen, because Telegram polls have a great many settings.
What the card shows you
- Question
- What is being asked, with its formatting.
- Options
- The answers people can pick from.
- Multiple answers
- Whether one person may choose more than one.
- Quiz mode
- Turns the poll into a quiz with one correct answer, and names it.
- Anonymous
- Whether votes are attributed to whoever cast them.
- Open period
- How long the poll accepts votes before it closes.
- Revoting
- Whether somebody can change their mind after voting.
- Shuffle options
- Whether the options are shown in a different order to each person.
- Hide results
- Whether the running total stays hidden until the poll closes.
- Adding options
- Whether voters may add options of their own.
When you tap it
Opens the recipient picker, then creates the poll in the chats you chose.
You choose who, after you have read it
Tapping the card opens a multi-select recipient picker — send to one chat, or fan out to many at once. The card then shows its real-time status: idle, sending, sent with a recipient count, or failed. A partial send reads “Sent to 2 of 3 chats” and names the first error, and Retry pre-selects only the recipients that failed, so nobody is messaged twice.
See also
Watch out for
Everything below Multiple answers is Telegram-only. WhatsApp ignores those settings rather than refusing the poll, so the same card sent to a WhatsApp chat degrades to a plain question with its options — which is usually what you wanted, but it is not what the card shows you.
Event
What it is
An event lifted out of a conversation — the dinner buried across ninety messages — ready to go into your calendar. This is the card the Analyze button produces most often.
What the card shows you
- Title
- What the event is called.
- When
- Its start and end.
- Where
- The location, when the conversation gave one.
- Notes
- Anything else worth keeping with it.
- From
- The message it was read out of, so you can check it against what was actually said.
When you tap it
Opens the calendar editor with everything filled in. You get one more look — and a chance to change the calendar, add an alert or invite somebody — before it is saved.
Needs Calendar access
The tap writes into Calendar, so the first time you use this card iOS puts up its own permission prompt. Decline and the card simply cannot file anything — nothing else about the conversation changes.
Why this one opens an editor
Every other device card writes on a single tap. An event is the one thing people almost always want to adjust before it lands — which calendar, which alert, who else is coming — so it deliberately routes through the calendar editor rather than being created behind your back.
See also
Reminder
What it is
The thing you promised, turned into a task in Reminders before it is lost again.
What the card shows you
- Title
- What has to be done.
- Notes
- The detail behind it.
- Subtasks
- The steps it breaks into, when it has any.
- Due
- When it is due.
- Priority
- How urgent it is.
- List
- Which of your Reminders lists it belongs in.
- From
- The message it was read out of.
When you tap it
Creates the reminder in Reminders, subtasks and all.
Needs Reminders access
The tap writes into Reminders, so the first time you use this card iOS puts up its own permission prompt. Decline and the card simply cannot file anything — nothing else about the conversation changes.
See also
Action Item
What it is
Not a task you set yourself — a commitment somebody made, pulled out of the thread with the person who made it. This is what a group chat is full of and what a summary normally throws away.
What the card shows you
- Action
- What was promised.
- Owner
- Who promised it.
- Deadline
- When it is due — overdue ones are highlighted.
- Status
- Pending, in progress, completed or delegated.
- Blockers
- What is standing in its way, when anything is.
- Context
- The surrounding detail that makes the action make sense.
- From
- The message it was read out of.
When you tap it
Saves the action item and syncs it into Reminders. Tapping it again updates that same reminder rather than adding a second one, so an item you revisit as its status changes stays one row in Reminders.
Needs Reminders access
The tap writes into Reminders, so the first time you use this card iOS puts up its own permission prompt. Decline and the card simply cannot file anything — nothing else about the conversation changes.
See also
Contact
What it is
The new number somebody sent you, ready to become a real contact instead of a message you will have to search for later.
What the card shows you
- Name
- Who it is.
- Reaching them
- Phone number and email address.
- Where
- A postal address, when one was given.
- Work
- Company, job title and website.
- Notes
- Anything else worth keeping on the card.
- From
- The message it was read out of.
When you tap it
Creates the contact — or updates the one you already have, rather than making a duplicate.
Needs Contacts access
The tap writes into Contacts, so the first time you use this card iOS puts up its own permission prompt. Decline and the card simply cannot file anything — nothing else about the conversation changes.
See also
Note
What it is
Something worth writing down that is not a task and not an appointment — the reference, the address, the list somebody read out.
What the card shows you
- Title
- What the note is called.
- Body
- The note itself.
When you tap it
Saves it into Apple Notes.
See also
Image
What it is
A picture Apple Intelligence drew, on your iPhone, from a prompt — through Image Playground, in its Animation, Illustration or Sketch style.
What the card shows you
- Prompt
- What it was asked to draw.
- Image
- The picture itself.
When you tap it
Saves it to Photos, or hands it to the share sheet.
Needs Photos access
The tap writes into Photos, so the first time you use this card iOS puts up its own permission prompt. Decline and the card simply cannot file anything — nothing else about the conversation changes.
Drawn on the device
Image Playground runs on your iPhone, so the prompt and the picture both stay there. This is the one card whose content Claude did not write and did not receive.
See also
Link
What it is
A page worth opening — a source Claude cited, a place it found, a document it referenced.
What the card shows you
- Title
- What the page is.
- Address
- The URL it opens.
- Snippet
- A line of what is on the other end, so you know before you tap.
When you tap it
Opens the URL.
See also
Skill File
What it is
A finished document Claude built by running code in a sandboxed container — a spreadsheet with the numbers in it, a deck, a report — downloaded to your device and waiting to be looked at.
What the card shows you
- File
- The document, with its name, size and type: .pptx, .xlsx, .pdf or .docx.
When you tap it
Opens a full-screen preview. Share exports it wherever you want it — including into a chat.
See also
Watch out for
This card needs your own Anthropic API key. The document is produced in Anthropic’s code-execution container, which is not part of what prepaid AI Credit covers — on AI Credit alone the agent has everything else, but not this.
Question
What it is
The one card that is not an ending. Rather than guess at something it cannot settle on its own, the agent asks you — and your answer goes back in as your next message, so the work carries on from there.
What the card shows you
- Question
- What is being asked, under a short label naming the decision.
- Options
- The answers to choose between, each with a line explaining what picking it would mean.
- Preview
- On some options, a worked example of the result — a mock-up, a snippet, a variant — shown beside the choice.
- Other
- A free-text field, always offered, for an answer that is not on the list.
- Note
- Room to add a remark alongside whatever you picked.
When you tap it
Submit sends your answer back into the conversation as a new message from you, and the agent picks up from there with the answer in context. More than one question can ride on a single card, and a question can accept several answers at once rather than just one.
It stays where it was asked
Once answered, the card holds your answer read-only in the thread, so the reason the work went the way it did is still visible later rather than being a decision you made and forgot.
The agent chat — questions, answered.
The context window is all the text a model can reference when generating a response, including the response itself — a working memory for the model. A larger window handles longer, more complex prompts, but more context isn't automatically better, so AI Reply Assistant actively manages it for you:
- Compaction — automatically summarises older context as you near the limit, keeping the active context focused and performant
- Prompt caching — resumes from cached prefixes, significantly cutting processing time and cost on repetitive tasks
- On-demand tool loading — tool definitions load only when needed rather than all at once, preserving working memory
- Attachment deduplication — avoids re-sending the same file
When a conversation approaches its context limit, AI Reply Assistant prompts you to compact or start a fresh topic.
This describes the Claude path, where the window runs to 1M tokens on Sonnet 5.5 and Opus 5.5. Apple Intelligence works in a much smaller window on your iPhone — Private Cloud Compute is the tier to switch to when a turn needs more room than that.
MCP (Model Context Protocol) is an open standard for connecting AI assistants to the systems where data lives — content repositories, business tools, and development environments — replacing fragmented integrations with a single protocol. Think of MCP like a USB-C port for AI: just as USB-C standardises how you connect electronic devices, MCP standardises how AI applications connect to external systems.
MCP is designed primarily for desktop and server environments. AI Reply Assistant takes a complementary approach: instead of connecting to external MCP servers, the agent runs natively on iOS and executes tools directly through Apple's own frameworks (EventKit, Contacts, and others) inside the iOS app sandbox.
The result: your data benefits from iOS-level permissions, every action requires your explicit approval, and tools run with the same security guarantees as any native app.
Extended Thinking gives Claude enhanced reasoning for complex tasks, with visibility into its step-by-step thought process before the final answer. It is not a different model — it lets the very same model give itself more time and expend more effort to reach an answer.
You set the effort level to control how much time Claude spends on a problem — low, medium, high, or extra high; higher effort can improve quality through more thorough analysis. Opus and Sonnet use adaptive thinking that automatically adjusts effort to the task; Haiku uses a fixed budget.
A collapsible "Thinking" block with an elapsed timer shows a summary of the reasoning as it happens, making Claude's answers easier to understand and check.
When Claude searches the web or reads one of your documents, superscript markers appear in the answer where it drew on something. Tap the Sources pill under the response for the full list — favicon, title, domain, and the passage that was actually cited — and tap a web source to open the page.
The pointers come back with the response rather than being written into the text by the model, so a citation always resolves to a real passage in a real document.
Each completed response has action buttons: Copy (to clipboard), Share (via the iOS share sheet), and Speak (reads the response aloud, with a stop toggle). In overlay mode, a "Use as Reply" button lets you send the response directly back to your messaging conversation. You can also select any text natively for Look Up, Translate, or Writing Tools. None of this depends on which model produced the answer.
The agent chat opened from inside a conversation rather than from the tab bar — which is what the Skill button above the keyboard does. It arrives pre-scoped to that one chat and pre-loaded with the Skill you picked, and Use as Reply drops the answer into that chat’s composer.
The difference from a standing agent chat is scope and starting point, not capability: it is the same chat, already pointed at the thread you were reading.
Yes. Multi-agent orchestration lets one agent coordinate with others to complete complex work. Agents can act in parallel with their own isolated context, which helps improve output quality and time to completion. For example, one sub-agent can search the web while another analyses your messages. Each sub-agent has its own conversation history and restricted tool access — no message sending or further delegation. Only one level of delegation is supported: the coordinator can call other agents, but those agents cannot call agents of their own. Sub-agents are capped at 15 turns per task.
The agent loop runs on Anthropic’s servers instead of on your iPhone. Claude keeps the conversation history and directs its own multi-step work there, while your device still executes the tools locally and streams the results back — so a long job survives your phone locking, the app being backgrounded, or you picking it up on another device.
Toggle it per topic from the model picker. It needs Opus or Sonnet and your own Anthropic key: server sessions are organisation-scoped, so they are not offered on AI Credit.
Yes. The agent supports VoiceOver with labels and hints on every interactive element, Dynamic Type for scalable text, native text selection with Look Up, Translate, and Writing Tools, and haptic feedback on key actions. It also adapts fully to dark mode.
Your personal AI assistant for messaging.
Download AI Reply Assistant on iOS and try it free.