The AI keyboard for Tinder.

Tinder is the default inbox, which is why opener conversion is the real product

Pew Research Center's 2023 online-dating reports found that 14% of all U.S. adults have used Tinder, 42% of adults ages 18–29 have used it, and 79% of online dating users under 30 have used Tinder. That concentration is the operating fact I designed KOPY around: if you ship an iOS reply keyboard for dating, Tinder is the highest-volume first-message surface, not a niche swipe toy.

I did not build KOPY as a Tinder bot. I built a custom keyboard that sits in Tinder's composer because that is where the conversion event happens. Match volume is cheap. The scarce resource is a first send that earns a reply before the thread dies in the stack. When I instrumented my own Tinder usage while dogfooding the extension, unmatched likes barely mattered. Unanswered first messages did.

If you are comparing apps, Hinge prompt replies and Bumble first-move timers punish different failure modes. Tinder punishes delay plus generic openers. The research below is for that job: first-message conversion and chat decay after the match.

Pew's under-30 split explains why "just be funnier" is a bad KPI

Among U.S. adults under 30, Pew found a near-even split: 35% say dating apps make it easier to find a long-term partner and 33% say they make it harder. Tinder lives inside that ambivalence. People keep matching while distrusting the funnel. A witty line that could have been sent to anyone confirms the "harder" camp. A specific line that uses one photo, one bio fragment, or one location cue is the only opener that fights the 33%.

I treat Pew's split as a product constraint. Dating mode in KOPY is not a joke generator. It is a specificity engine with Rizz, Funny, Spicy, Agree, and Disagree as levers. Funny is for when the profile already asked to be roasted. Rizz is for when the photo is the hook. Disagree is for a bio take you can push on without being cruel. If the model returns a line that would work on the next fifty profiles, I rewrite it before I insert it. The keyboard's job is a draft in under two seconds. The user's job is one proper noun.

The Tinder KPI table I actually track

Pew's age and platform numbers tell you where volume lives. They do not tell you whether your Tinder account is converting. I refuse to publish fake KOPY user counts or App Store star averages. Track these four metrics on your own matches for two weeks. Write them in Notes. Ignore super-like vanity.

KPI	How to count it	What "good" looks like in practice
Time-to-first-reply	Minutes from match timestamp to your first send, during waking hours	Under 30 minutes on days you are actually using the app; same-day if you batch at lunch and 9pm
Opener reply rate	Incoming replies divided by first messages you sent	Track the weekly rate, not one lucky night. Generic "hey" tanks this; photo-specific openers lift it
Continuation past 5 messages	Threads where you and they combine for five or more bubbles	This is the quality filter. A reply that dies at "haha" is a failed opener wearing a costume
Date-ask rate	Explicit plan asks (day + neighborhood + activity) divided by threads that cleared 5 messages	Ask after you have a hook, usually messages 6–8. "We should hang" is not a date ask

Chat decay is the silent fifth metric: matches with zero messages after 24 hours. On Tinder that pile is huge because matching is frictionless. Shrink decay before you optimize date-ask rate. You cannot ask people out from empty threads.

Why I shipped an iOS keyboard instead of a Tinder overlay

Apple's custom keyboard documentation is the reason KOPY exists as an extension, not as a floating assistant. Third-party keyboards run in a tight sandbox: roughly 70MB of RAM, no background URL sessions, and network plus pasteboard access only when the user grants Full Access. I learned this the hard way in the first prototype. Without Full Access, the system pasteboard is empty and the model has nothing to reply to. With Full Access, the user copies the incoming line (or a bio snippet), switches to KOPY with the globe key, and we read clipboard context.

Generation uses an ephemeral URLSession because keyboard extensions are not allowed to pretend they are background download daemons. The request is a short POST with mode, tone, and context. The product promise I held myself to is a draft in under two seconds on a normal LTE connection. If the call is slower than that, people bounce back to the system keyboard and type "hey" because Tinder's chat list is already scrolling past the match.

That architecture is why KOPY can live inside Tinder's text field instead of forcing a round-trip through Notes or a web chat UI. Tinder never has to know we exist. There is no AI badge. The inserted text is ordinary keyboard input. You edit it. You send it. If a line feels templated, delete it. The extension is a first draft with a RAM budget, not a personality transplant.

Opener research: the conversion unit is the first bubble, not the bio

Pew's finding that Tinder reaches 42% of 18–29-year-olds means your opener competes with a lot of other first bubbles the same evening. I ran a personal n of a few hundred first messages while building Dating mode. The pattern was ugly and consistent. Openers that referenced a specific photo detail (dog breed, hike location, jersey, book spine) earned replies. Openers that announced my personality ("I'm sarcastic but sweet") did not. Openers that asked two questions stacked like an interview did not. Openers that copied a meme format everyone else was using that week died in the same hour they were born.

The practical rule: one observation plus one easy turn. Observation from their profile. Turn they can answer in a sentence. Example shape: "That alpine lake looks fake in the best way. Was that a one-hour hike or a full-day sufferfest?" That line has a proper noun, a compliment that is not about their face, and a binary they can riff on. KOPY's Rizz tone tends to over-index on swagger. I keep Rizz when the photo is already flirtatious. I switch to Funny when their bio is already joking. I switch to Insight only when the profile is actually saying something (travel, books, a take), because Insight mode on a mirror selfie is how you get a TED Talk nobody asked for.

We published a longer opener teardown at why your Tinder openers aren't getting replies. This page is the KPI and keyboard version of that argument: speed plus specificity, measured, not vibes.

Chat decay after the match: Tinder's second death

Even among people who use apps, Pew's 35/33 split under 30 is a reminder that many users experience the products as noisy. Tinder accelerates that noise. A match is a weak signal. If the first reply takes six hours, the other person has already stacked new matches on top of yours. Decay is not mysterious. It is a queueing problem.

I batch Tinder twice a day on purpose: a short pass after lunch, a short pass after dinner. Each pass is copy, globe, Dating mode, generate, edit, send. I do not "live in the app." Living in the app produces anxious micro-edits and late-night messages that read thirsty. The keyboard's two-second draft is what makes batching possible. Without it I stall on message three of a dying thread because "lol" has no handle.

Continuation past five messages is where most people fake progress. They count any reply as a win. I count a thread as alive only when both people have asked something and answered something. If they send "haha true" I do not send another joke. I send a Continue-shaped move: a fork they can pick. "Two paths: you tell me the worst concert you've survived, or we skip to coffee near [neighborhood]." Continue as a tone lives in Insight mode on the keyboard; in Dating mode I approximate it with Funny plus a concrete ask, or I flip to Social mode if the chat has already left flirt-land and become weekend logistics.

Keyboard workflow I use inside Tinder (not a generic how-to)

Apple requires Full Access if a custom keyboard needs the network and a useful pasteboard. That is not marketing language. It is the permission dialog you will see. I walk new testers through it because people assume a keyboard can "see Tinder." It cannot. Tinder's web view does not hand us the last message. You copy. We read the copy. We generate. You insert.

1. Copy the line you are answering. If the chat is empty, copy one bio line or a photo caption you can actually see.
2. Tap the globe until KOPY is the keyboard. Confirm Dating mode. Pick Rizz, Funny, or Spicy based on the hook, not based on your mood.
3. Generate. Wait for the sub-two-second draft. If the draft has no proper noun from their profile, add one before you send.
4. If the thread is already three messages of "haha," switch the tone and generate again instead of stacking another pun.
5. When you are ready to leave the app, ask for a plan with a daypart and a place type. Do not ask "what are you looking for" in message two.

Work mode and Insight mode are not the Tinder default. I use Work if someone is clearly talking about a job change and I want to sound like an adult. I use Insight if they dropped a genuine question about a book or a city. Social mode is for the post-date logistics text that still happens inside Tinder for people who have not moved to iMessage yet. Dating mode remains the conversion engine for the first five bubbles.

What I will not do in Tinder, even though the model can

Pew documented broad Tinder reach among young adults. Broad reach is not a license to spray. I do not generate sexual first messages to people who did not invite that register. Spicy is a later-thread tone when they have already gone there. I do not generate "I hate games, let's be real" because that sentence is a game. I do not generate long voice-note transcripts as openers. Tinder's UI rewards scannable lines. A 90-word paragraph looks like homework.

I also do not let the keyboard invent facts about their life. If the clipboard only contains "hey," the model should not pretend it saw a golden retriever. Copy better context or send a photo-based observation you typed with your eyes. The clipboard-plus-Full-Access design is honest about what the model can see. Anything else is a hallucination with a dating cost.

Where Tinder sits next to Hinge and Bumble

Pew's stack among 18–29-year-olds is Tinder at 42%, Bumble at 27%, Hinge at 21%. Tinder is the volume surface. Hinge is the prompt-comment surface. Bumble is the timer surface. If your Tinder opener reply rate is fine but you never get dates, you have a continuation problem, not a swipe problem. If your continuation is fine but you never ask, you have a date-ask problem. KOPY can draft the ask. It cannot want the date for you.

Use this page when your search intent is Tinder first messages and decay. Use the Hinge page when the unit of work is a prompt. Use the Bumble page when the unit of work is a 24-hour first move. The keyboard is the same process. The KPIs are not.