AI reply
Draft an answer to a message somebody sent you: paste it, type it, or pick a screenshot of the chat, then choose a tone.
The AI tool works on text you have already written. AI reply works the other way round: it reads what somebody else sent you and drafts answers for you to pick from. It is for the moments when you know what you want to say and not how to say it, or when you do not know what to say at all.
Nothing is ever sent on your behalf. The tool writes drafts; tapping one puts it in the message box, exactly as if you had typed it, and you still press send yourself.
Getting their message in
Section titled “Getting their message in”A keyboard can read the box you are typing in. It cannot read the conversation above it, because those messages belong to the chat app, and no keyboard can reach into them. So the first thing the panel asks for is their message, and there are four ways to hand it over:
- Copy it. Long-press their message in the chat app and copy it. Anything you copied in the two minutes before opening the panel is already there when it opens, and anything you copy while it’s open is added as you copy it. This works whether or not the clipboard keeps a history; without one, only the latest copy is picked up on opening. Paste copied adds the most recent copy by hand.
- Screenshot shows the recent screenshots on your phone as a grid of pictures, newest first, each marked with how long ago it was taken. Tap the one of the chat.
- Type opens a box on the keyboard. Type or paste the message, press Enter, and it is added.
- Use the selection works in the few apps that let you select the conversation text itself.
Several messages at once
Section titled “Several messages at once”A reply usually needs more than one line of the conversation. While the reply panel is open, anything you copy is added to it automatically, so you can switch to the chat app, copy three messages one after another, come back, and all three are waiting. Screenshots can be picked up the same way; see below.
The gathered messages are drawn like the chat they came from: theirs on the left, yours on the right in the theme’s accent colour. Each carries a Them or Me label; tap the label to flip it. This matters more than it looks: nothing the keyboard can see says who wrote a copied line, so a message of your own that is read as theirs produces a reply to yourself.
Tap a message’s text to edit it in the box on the keyboard; Save puts it back, and saving it empty removes it. Swipe a message sideways, or tap its ×, to drop it, or tap Clear to start over. Between two messages sits a small Join button: when one message arrived as two pieces, tap it to make them one again.
A message with no real word in it is outlined in red and marked No real words?. That is usually what reading a screenshot makes of a sticker, a profile picture or a line cut off at the edge. Only words in Latin letters are judged, against the dictionaries your keyboard has loaded. Messages in other scripts, emoji and numbers are never marked. Remove unclear at the front of the row clears them all at once.
A screenshot sent as a picture has no Them or Me label. It holds both sides of the conversation, and the model reads who said what from where each bubble sits.
Screenshots
Section titled “Screenshots”A screenshot is the fastest way to hand over a whole conversation, because the bubbles’ own left-and-right alignment says who said what.
How it is read depends on your AI provider, set under WM Keyboard Tools AI reply How screenshots are read:
- Automatic (the default) sends the picture to providers that accept pictures (Claude, OpenAI, Gemini, xAI, Ollama, LM Studio and OpenAI-compatible servers) and reads it on the device otherwise. With an on-device model, Gemma 4 E2B and E4B look at the picture themselves, without it leaving the phone. A model you imported is tried the same way; if it turns out not to read pictures, the screenshot is read on the phone instead and the drafts are written from that, with a note in the panel saying so.
- Send the picture always sends it. The model reads the alignment itself, so no labelling is needed.
- Read on this device never sends the picture. The text is read on the phone and arrives as separate messages you can relabel. Full edition only: the on-device reader is not in the Lite build.
Once a screenshot is in the conversation, a Screenshot goes as row above the tones switches it between Picture and Text for this reply. Text reads it on the phone straight away, so you can check and relabel every line. Picture puts the lines read from that screenshot back together into the picture. Picture only shows when the chosen model can take one. Your choice is remembered as the setting above: Picture as Automatic, so switching later to a model that can’t read pictures falls back to text instead of failing, and Text as Read on this device. If you switch to a model that can’t read pictures after adding one, the screenshot is read on the phone when you pick a tone.
When a screenshot is read on the device, the parts that are not the conversation are left out: the status bar and the chat app’s top bar (the contact’s name, “Active now”), message times and dates, “Seen” and “Delivered” marks, the message box’s own placeholder, and the app’s own notices inside the thread, such as “Replied to you”, reaction banners, “This message was deleted”, “Photo” and “Voice message” labels, call records, unread and encryption notices, and group events. A screenshot taken while the keyboard was on screen has the keyboard in it, and everything from the keyboard’s top edge down is skipped. The keyboard remembers where its top edge was at each moment, so a screenshot taken with the plain keyboard up is cut at the plain keyboard’s top, even when you pick it later from the taller reply panel. Lines of one bubble that wraps are put back together; when that misses one, use Join.
The keyboard cannot take a screenshot for you: capturing the screen needs a system permission dialog, and an input method that opens one loses the box it was typing into. Take it the usual way, with the hardware buttons, and pick it with Screenshot. To skip that step, turn on WM Keyboard Tools AI reply Add screenshots automatically. It’s off by default. With it on, a screenshot taken while the panel is open is added on its own, and so is the newest one from the two minutes before you open it. This works whatever Screenshots in the clipboard under WM Keyboard Clipboard is set to, and a screenshot caught for a reply is not added to the clipboard unless that setting is on. It needs the same permission to read your images as the Screenshot button.
Reading screenshots from the gallery needs permission to read your images. Tapping Screenshot without it asks for it straight away, and the grid opens once you allow it. If you said no, the panel shows an Allow button to ask again.
Picking a tone
Section titled “Picking a tone”Once there is something to reply to, the tones appear at the bottom of the panel under Next: pick how to reply, in one row you can swipe sideways. They stay there however long the conversation gets, and the conversation above them scrolls to the newest message as you add one. One tap writes the drafts:
| Tone | What it does |
|---|---|
| Casual | A relaxed text, the way friends message each other |
| Quick | Two to six words, or a single emoji when it fits |
| Keep it going | Answers, then asks one easy question back so the chat continues |
| Sounds good | Says yes and confirms any plan, time or place |
| Say no | Declines clearly and kindly, without repeated apologies |
| Funny | Playful banter that still answers what they said |
| Excited | Hyped or congratulating, the way a friend texts back |
| Supportive | Shows you understood how they feel, without unasked-for advice |
| Apologise | A short, sincere apology without excuses |
| Polite | Courteous and professional, for someone you do not know well |
| Flirty | Warm and lightly flirty |
| Custom… | Asks what the reply should say, then writes it in that direction |
Custom… turns the box on the keyboard into an instruction box: type something like “say I’m running late, suggest 8:30” and tap Write.
Tones are editable, and you can write your own under WM Keyboard Tools AI reply Tones. Reorder them, turn some off, change what one does, or add a new one with New tone. A tone you write can also ask for its instruction each time, the way Custom… does. Changing a shipped tone keeps a Reset to bring the original back.
The drafts take the place of the conversation, and each one appears as soon as the model has finished writing it, rather than all three at the end. Tap one to put it in the message box.
Starting from your own words
Section titled “Starting from your own words”If you have already typed something in the message box when you open the panel, such as “cant fri, busy”, the drafts say that, in the tone you pick, instead of inventing a different answer. The panel shows From your draft above the tones so you know it is being used. The draft you pick replaces what you typed. If you typed more in the meantime, it is added at the cursor instead and your text is left alone. The tone they were written in stays highlighted; tap another under Try another tone for a different set, or Messages to go back and fix a message first.
What it writes
Section titled “What it writes”Under WM Keyboard Tools AI reply:
- Replies per run: how many alternatives one run writes, up to five. A small on-device model is asked for one, because asked for three it writes the same reply three times. The recommended on-device models, Gemma 4 E2B and E4B, get the full number.
- Length: short, medium or long.
- Emoji: match the other person’s use of them, allow them, or avoid them entirely.
- About you: anything the drafts should know about how you write. This is the single setting that most changes whether a reply sounds like you.
- Reply language: blank follows the conversation, including its script, so Bengali written in Latin letters is answered the same way.
- Conversation sent: how much of the gathered conversation one run sends. When it will not all fit, the oldest messages go first.
The provider, API key and model are the AI tool’s own, under WM Keyboard Tools AI. There is only one set of them. When more than one is set up, the panel shows the same Model row as the AI tool at the top, with your downloaded on-device models and every cloud provider you have a key for. Picking one there picks it for the AI tool too. The row is hidden while drafts are being written, and turns off with Model picker on the panel under WM Keyboard Tools AI.
Clearing up afterwards
Section titled “Clearing up afterwards”Feeding this tool leaves other people’s messages in your clipboard history and screenshots of your conversations in your gallery. So when you tap a draft, the copies and screenshots that reply was built from are taken back out again. It is on by default, it is a checkbox in the panel if you want to keep one for a particular reply, and it is a setting under WM Keyboard Tools AI reply Clear what it used.
Clipboard entries are the keyboard’s own, so those always go. A screenshot in your gallery is yours: from Android 11 an app may only delete pictures it created itself, so Android shows its own dialog asking whether WM Keyboard may delete it. The reply is already in the message box by then, so answering the dialog loses nothing. On Android 10 and older, or when the dialog can’t be shown, the panel says plainly that the screenshot is still in your gallery.
Privacy
Section titled “Privacy”- Nothing leaves the device until you tap a tone. Gathering messages, labelling them and reading a screenshot on the device are all local.
- Their message is treated as data, never as instructions. A line in a conversation that tells the model to ignore its instructions is answered as part of the conversation, which is what someone sending it would deserve.
- The tool refuses to run in a password field.
- Drafted replies are not kept in the AI history, even when the history is on. A record of a rewrite is your own writing; a record of a reply is somebody else’s message. Turn it on under WM Keyboard Tools AI reply Keep drafted replies in the AI history if you want them kept.
- On a metered connection, data saving holds the request the same way it holds any other cloud AI run.
- Choosing an on-device model keeps the whole thing on the phone, screenshots included. Gemma 4 reads the picture itself; other on-device models get the text read off it on the phone.
- The draft you had started in the message box is sent along with the conversation, since the drafts are written from it.