The AI keyboard for WhatsApp replies.

WhatsApp is not a niche messenger you dip into. Meta describes the WhatsApp family of apps as serving more than two billion users. That scale is the point: family logistics, neighborhood groups, school-parent threads, and mixed work chats all live in one inbox. The bottleneck is not typing speed. It is deciding how to sound, to whom, while the next ping arrives.

We built KOPY as an iPhone keyboard because WhatsApp already owns the conversation surface. Opening a separate assistant, pasting a screenshot, and copying a draft back is a context switch that kills household cadence. A keyboard stays in the thread. That is the product constraint, not a slogan.

What 2B+ users actually changes about your reply load

Headcount at Meta's scale does not mean your aunt is waiting on a corporate SLA. It means WhatsApp absorbed roles that used to live in SMS, email, Facebook groups, and hallway talk. Birthday planning, who is picking up the kids, and are you still coming Sunday all compete with a work ping that accidentally landed in the same app.

The failure mode is backlog-by-relationship. You answer the easy one-to-one first. Family groups with four overlapping threads wait. Emotionally loaded one-to-ones wait longer. Speed here is measured in minutes-to-hours of silence, not keystrokes per minute.

If you want a tactical walkthrough of iOS notification replies, Siri, and keyboard flow, use our field guide on how to reply to WhatsApp messages faster. This page is the systems view: household chat as a queue with mixed stakes.

Personal, family, and group chats are three different jobs

Treat WhatsApp as three products sharing a UI. Personal chats need warmth and specific callbacks. Family chats need logistics plus face-saving. Groups need brevity that does not perform for an audience.

• One-to-one personal: the cost of delay is relational. A late apology still works if the sentence names the original ask.
• Family: the cost of delay is coordination. Ambiguous yes or no answers create follow-up pings.
• Groups: the cost of delay is social. Over-long replies look like speeches. Silence looks like snubbing.

KOPY's Social mode exists for this mix. Funny is for group re-entry. Neutral is for I felt ignored. Disagree is for the relative who turns a plan into a character judgment. Continue is for the thread that died at lol ok. None of those jobs is Instagram-story wit or a public X ratio. They are household register.

A copy-then-generate loop that respects WhatsApp's UI

iOS custom keyboards cannot read the WhatsApp bubble directly. Apple's model is insert-text-at-cursor. So we take clipboard as context: you copy the incoming line, switch to KOPY with the globe key, pick mode and tone, generate, then edit in the composer. That is clumsy until it is muscle memory. After that it is faster than drafting from a blank field while the group keeps moving.

Apple documents that a keyboard without open access cannot use the network or UIPasteboard. Allow Full Access is the user-facing name for expanding that sandbox. See Apple's custom keyboard documentation and the App Extension Programming Guide. We cannot hide that toggle. Clipboard context is how generation stays tied to the actual message.

First-person constraints: 70MB, clipboard, no background magic

We ship inside Apple's keyboard extension budget. Roughly 70MB of RAM is the practical ceiling. That is why KOPY is not a miniature chat client with history, plugins, and a document store. We use an ephemeral URLSession, generate, and get out. We do not keep a WhatsApp archive on device for better memory. If we did, the extension would die under memory pressure and you would be staring at a blank keyboard in a family emergency thread.

Clipboard is a feature and a liability. Full Access lets the extension read pasteboard text so the model sees what you copied. It also means you should not copy a one-time password into the pasteboard and then open KOPY. Secure fields already force the system keyboard. Treat the pasteboard as hot context, not a vault.

Generation still needs a network call. Apple's open-access keyboard can send keystrokes and other data for server-side processing. That is the trust contract: you enable Full Access because a local-only keyboard cannot hit our generate endpoint. We process the copied snippet to draft a reply. We do not turn WhatsApp into a logged helpdesk transcript.

Household speed protocol we actually use

When we test WhatsApp internally, we do not optimize for literary quality. We optimize for whether this unblocks the next physical action. Pickup times, addresses, dietary constraints, and I'll call you in 20 beat a paragraph of vibe.

1. Copy the line that contains the ask, not the whole forty-message scrollback.
2. If the chat is family logistics, start in Social plus Neutral or Agree. Funny is a tax when someone needs a time.
3. If the chat is a roasting group, Funny or Continue. Do not dump Work-mode Professional into a cousins thread.
4. Edit one proper noun. If the draft missed the restaurant name, that is the only required human pass.
5. Send before the group produces three more messages that make your reply look late to a previous topic.

Group re-entry after weeks of silence is its own genre. People write novels. The thread wanted a one-liner that acknowledges absence without a lecture. I was in witness protection is a joke register. Sorry, overloaded, I'm here now — still on for Saturday is a logistics register. Pick one. Mixing both is how you get a speech.

Tone switching without leaving the thread

The same phone handles a partner's long voice-note transcript, a parent's guilt text, and a building WhatsApp about package theft. Manually retuning voice across those in two minutes is the overthinking tax. Mode plus tone is a constraint set, not a personality transplant.

Work-adjacent WhatsApp — client groups, contractor threads — should not impersonate Slack. People still expect slightly human punctuation. Professional tone in Work mode is for yes I can review by end of day. Insight mode is rarely the right WhatsApp tool unless the group is actually debating a plan. Dating mode stays out of family chats. That sounds obvious until you leave the last keyboard state on Rizz and reply to your uncle.

Emotionally loaded one-to-ones are not faster-equals-better

Speed is the wrong metric when someone said they felt ignored. The job is to answer at all, with a sentence that names the feeling and a next step. Neutral exists for that. Generating a Funny draft and sending it because it appeared first is how you create a second conflict. Review the first clause. If it jokes about the injury, rewrite or switch tone.

KOPY will still produce a first draft in a couple of seconds. Your job is the accept-or-reject gate. Keyboard AI is a stub, not a diplomat.

KPI table: WhatsApp jobs, not vanity speed

These are design KPIs we use when judging drafts. They are not survey results and not a claim about your personal latency.

Chat type	Primary job	Failure if slow	Draft KPI we score
One-to-one personal	Acknowledge plus specific callback	Relationship cools	Names the ask in the first sentence
Family planning	Time, place, owner	Three follow-up pings	Contains a concrete when and where
Family tension	Face-saving repair	Escalation overnight	No joke in sentence one
Friend group	Re-enter without a speech	You look checked out	Under two sentences, one hook
Mixed work-on-WhatsApp	Clear next action	Thread forks	Owner plus deadline, no emoji pile

If a generated line fails the KPI for that chat type, switch tone rather than adding adjectives. WhatsApp groups punish ornate language. They reward a send that lets other people stop waiting on you.

What we refuse to pretend

We will not invent an average WhatsApp user sends forty-seven messages a day, or a conversion rate for family harmony. Meta's public claim is user scale, not your household's ping rate. We also will not claim KOPY reads WhatsApp's servers. Clipboard plus keyboard insertion is the whole bridge.

We will not claim Full Access is optional if you want clipboard-conditioned drafts. Apple's extension model is explicit: pasteboard access sits behind open access. If you decline Full Access, you can still type. You cannot honestly feed the last message into generation without some context channel. Copy is that channel.

For setup, pair this page with how to enable a custom keyboard on iPhone and the WhatsApp-specific speed guide linked above. For overthinking that is not WhatsApp-specific, the introvert and ADHD pages are adjacent jobs, not the same bottleneck as a forty-person family group.

A realistic day in a two-billion-user inbox

Morning: school group, three photos of a form you already signed. The useful reply is got it, I'll send the scan after drop-off, not a caption contest. Midday: partner asks if you can move dinner. That is Neutral plus a counter-time. Evening: cousins group quotes a voice note. Funny, short. Night: a parent message that is half logistics, half scorekeeping. That one you do not speed-run with Continue. You copy the charged sentence, generate Neutral, and you keep the apology if it is true.

The keyboard does not know your family politics. It knows the copied text, the mode, and the tone. That is enough to beat a blank composer when you are holding groceries. It is not enough to replace judgment in the scorekeeping thread. Use it as a stub generator for the first class of messages, and as a phrasing aid for the second. Household WhatsApp is high frequency, mixed stakes, and allergic to essays. Write like you are unblocking a kitchen, not publishing a newsletter.

Voice notes create a second speed problem. Someone talks for ninety seconds. You get a transcript blob on the pasteboard if you copy it, or nothing if you only listened. Copy a sentence that contains the ask, not the whole monologue. Generate against that sentence. If there was no ask, the job is acknowledgment plus a question that is easy to answer by voice later.

Status replies and disappearing messages tighten the window further. A status is closer to Instagram than to a family planning thread, but it still lives in WhatsApp. Keep those to one concrete noun. Do not import Insight-mode analysis into a sunset photo from your cousin.

Broadcast lists and large communities amplify the group-speech problem. If you are not an admin posting rules, your reply should not read like a policy memo. Social plus Continue, then cut. Admins can use Work plus Professional for pin-worthy updates, then strip corporate filler so it still sounds like a person who lives in the building.

If you only change one habit: copy the ask, not the vibe. Generate against the ask. Edit the proper noun. Send. The two-billion-user network will produce another message whether or not your prose was perfect. Your family needed a time, a yes, or a repair — usually in that order.