I am one of the two people who make Subtext, so weigh my stake accordingly. Every claim below links to where it comes from, most of them to the app site itself. Fifty questions, in the order people tend to ask them.
A note on the numbers. Downloads, ratings and review quotes are the ones published on subtext.it, dated where the site dates them. I did not round anything up.
Why it exists
Why did you build Subtext?
Because of the message that gets typed, read back, and deleted. Someone writes the thing they actually mean, looks at it, decides the wording carries more risk than the point, and sends nothing, or sends a flatter version and regrets that instead. I kept seeing that pattern in my own threads and in other people's. Subtext exists so that the draft gets read once before it goes, by something that will tell you how it lands without being polite about it.
Was there one message that started it?
No single one, and I would be inventing a story if I named one. It was the accumulation. The apology that quietly turned into a defence. The reply of two lines that took an hour. The cancelled plan followed by twenty minutes of spiralling about the wording. In one of the articles on the app site I call that deletion the actual subject, and that is still the most accurate way I have of saying where it came from.
Why do people delete messages instead of sending them?
Because the sentence is harder than the argument. In most hard messages the person already knows what they think. What they cannot judge is how it will land, and the cost of getting that wrong attaches to them personally. So the message stays in the notes app. The research on avoiding replies frames it as emotion regulation, not time management. Sirois and Pychyl made that case for procrastination in general in a 2013 review, and I have written up how far it carries to messages, with the parts that are inference marked as such.
Sources Sirois and Pychyl, 2013, Procrastination and the priority of short-term mood regulation
Why can I not hear how my own message sounds?
Because you hear the tone you intended while you write, and you cannot switch that off. That makes people reliably overconfident about irony, teasing and the line between firm and cold, and reasonably accurate about plain warmth. The reader gets none of your intention, only the words. The classic study is Kruger and colleagues in 2005, who found people badly overestimate how well email carries their sarcasm. Two cautions, both in my notes: that study has never been replicated at high power, and a 2025 study found readers get the basic tone, positive or negative, of ordinary messages about right. The misreads live in irony and in the line between firm and cold.
Sources Kruger, Epley, Parker and Ng, 2005, Egocentrism over e-mail · Pollmann and Roos, 2025, People correctly interpret the tone of text messages and emails
Where does the feeling in a text come from, the words or the reader?
Mostly the reader, and that changed how I think about the whole product. The sharpest evidence comes from people with aphantasia, who cannot picture things and who report that feelings fade almost as soon as the trigger is gone. A thought that arrives as a picture hits harder than the same thought as information, which is Emily Holmes's 2008 argument, and Wicken, Keogh and Pearson tested it in 2021 by measuring how much people's palms sweat at frightening stories, and the aphantasic group's response was flat. The writer reads their draft as information. The reader builds the picture. That article is the one I went down the rabbit hole for.
Sources Holmes and Mathews, 2008, Mental imagery as an emotional amplifier · Wicken, Keogh and Pearson, 2021, Proceedings of the Royal Society B
Why have Apple and Google not built this into the phone?
Every phone now rewrites messages. Apple Writing Tools, Gboard and the chatbots will all make a draft friendlier or more professional, and Grammarly's tone detector will tell you how a sentence sounds. None of them reads the whole conversation and tells you how your reply is likely to land in it. I think that is deliberate. Telling someone their message sounds passive aggressive is a rude thing for a phone to do, and a platform cannot afford to be rude to a billion people. A tool you install on purpose can. The full comparison is here, including where the big ones beat us.
Sources Apple, Writing Tools in Apple Intelligence · Grammarly, tone detector
Why is it called Subtext?
Because that is the thing it reads. The site's own line is that every message is an iceberg. The words are the part above the water. What the message carries underneath, the guilt in a cancellation, the defence inside an apology, the hurt behind a short reply, is the subtext, and that is what the app names before it touches a word.
What does a messaging app have to do with being the Growth Therapist?
Same instinct, two surfaces. Who Samet Durgun is and what the consulting work covers comes down to reading an ad account before spending a cent in it. Subtext reads a message before you send it. The app site calls it a therapist for messages, and I did not plan the echo, it just turned out that I build the same tool twice. Even the writing rules match: disclose the stake first, link every study, cut the numbers that do not hold up.
Why did you write more than forty research articles for a messaging app?
Because for the first two years I explained the app badly. I would say it analyses the emotional tone of your draft, which is true and tells you nothing. So I sat down and read the actual research on emotion, tone and misreading, to find out whether the thing we built has a foundation under it. It mostly does, with a couple of cracks I did not expect. The articles are that reading, published with the sources so you can weigh my reading against them.
What it is
What is Subtext, in one sentence?
Subtext is an app that reads the emotional tone of a message before you send it, then helps you say the same thing in your own voice. The rest of this page is the longer version of that sentence.
Who makes Subtext?
Two people in Berlin. Anton Licht writes the code. I do everything that involves words, including the app copy, the articles on the site, and this page. There is no team behind us and no agency; the about page lists exactly the two of us.
Where is Subtext made?
Berlin, Germany. The legal entity behind it is registered there and the details are on the legal page, including the address, which I would rather you read there than have me paraphrase.
How long has Subtext existed?
About two years of building as of August 2026, which is when I wrote that figure into one of the articles. As of August 2026 the app site reports more than 100,000 downloads and a 4.9 rating on the App Store. Those numbers come from our own site, so weigh them accordingly, and treat that site as the record, not this page.
Is Subtext a chatbot?
No. A general chatbot writes a message for anyone. Subtext starts from a message you already wrote, or from a conversation you already had, and reads that specific exchange. It lives where you write, on your phone and in the browser, instead of in a separate window you have to go and prompt.
Is Subtext therapy?
No, and the app says so in its own fine print. It is a writing and communication assistant. It does not provide therapy, diagnosis or medical advice, its suggestions are not always right, and it will never know your relationship better than you do. Read anything before you send it.
Is Subtext only for texting?
No. People use it most for texts and hard personal conversations, and just as much for Slack messages, work email, feedback and the messages that chase a reply. Judging by the reviews, roughly a third of what it gets used for is work. If the message is written to another person, it applies.
What it does
What happens when I paste a message into Subtext?
It reads first and rewrites second, and the order matters. It marks the exact words that could land as harsh, cold or passive aggressive, names the feeling underneath them, and gives you a read on whether it is safe to send. Only then does it offer versions, side by side, and you pick one or steal the good parts and fix the draft yourself. The page that shows how it works plays one real message through the whole loop.
What kinds of triggers does it look for?
48 kinds, according to the app's own feature page. The ones that show up most in the real examples on the site are passive aggression, putting yourself down, sarcasm, loaded language and overpromising. Each tag explains itself, so you see why a phrase was flagged, not only that it was.
What does naming the feeling underneath mean?
Under the tags on the words, it names the emotion the draft is carrying: guilt, hurt, feeling overwhelmed, anxiety. It reads from a vocabulary of 51 emotions. That step is there because most of us work with about three words for what we feel. Brené Brown's team asked around 7,000 people over five years to name what they were feeling, and the average answer was three: happy, sad, angry. A hard message usually carries more than three. Naming it changes what you do about it.
Sources Brené Brown, Atlas of the Heart, on the survey of 7,000 people
What is the safe to send score?
An estimate from 5 to 100, attached to each version, of how likely the message is to land the way you mean it. Under it, one of two verdicts: safe to send, or check before sending. It is a read, not a guarantee, and the app is explicit that you know the relationship and it does not.
Why does it analyse before it rewrites?
Because naming what is inside a message changes how you handle it, and the evidence for that is decent. People who can tell their emotions apart in finer detail handle them better, a pattern Kashdan, Barrett and McKnight reviewed in 2015, and putting a feeling into words measurably lowers the brain's alarm response to it, which is Lieberman's 2007 finding. Most AI tools skip straight to the rewrite. We put the reading first on purpose, and the studies are linked here.
Sources Kashdan, Barrett and McKnight, 2015, Unpacking emotion differentiation · Lieberman and colleagues, 2007, Putting feelings into words
How many versions does it give me?
Three, side by side, each carrying its own tone label and its own verdict on whether it is safe to send. In the example on the site they are a warm and clear one, a reassuring one, and a short and kind one. You read them against each other and send whichever sounds like you.
Can I adjust a version without writing a prompt?
Yes. There are buttons for shorter, warmer, more casual, add humour and add confidence. Tap one and it rewrites while keeping your phrasing. Nobody has to learn to prompt.
Can it read a screenshot of a whole conversation?
Yes. Drop a screenshot of the thread and it reads the whole conversation, both sides, not only your line. That is what lets it tell you what the other person is actually asking for, separate from everything wrapped around it.
Can I talk to it instead of typing?
Yes. Record a voice note and talk the message out. It shapes what you said into something sendable. It also works on voice messages other people send you, so you do not have to listen to one that runs five minutes twice to find out what was being asked.
What does it mean that it thinks it through with you?
When a message really matters, it stops and asks what you want the reply to get across, then writes from your answer instead of guessing. In the site's example the choices are along the lines of "the small thing still matters to me" and "we are okay, no hard feelings". The version you get depends on which one you pick.
Does Subtext send the message for me?
No. Nothing is sent on your behalf. You copy the version you want and send it yourself, from wherever you were writing. That is a deliberate line. The effort a message is supposed to carry stays yours.
Whether it still sounds like you
Will the rewrite still sound like me?
That is what it is built for, and it is the most common theme in the reviews by a wide margin. It starts from what you already wrote instead of generating a message from nothing, and it shows how much it changed so you can pull it back. One App Store reviewer, quoted on the reviews page, describes finally being able to write in a way that feels like them, just warmer and clearer.
Why do AI rewrites usually sound robotic?
Because the button on your phone hands everyone the same smooth paste. Friendly, professional or concise, the output is polite in exactly the same way for every person who presses it. Subtext rewrites inside your own phrasing and rhythm, which is slower to build and the only version of this worth paying for.
Does using an AI to write a message make it land worse?
It can, and the mechanism is specific. In an experiment with 208 people imagining a close friend had used AI to write to them, the AI help made the message read as less effortful, and that lower perceived effort was what dragged down satisfaction with the friendship. Asking a human for help scored the same. The problem was never AI, it was outsourcing the effort. A tool that reads your own draft back to you sits on the right side of that line. A tool that writes it for you does not. The study is Liu, Kang and Wei, 2023, in the Journal of Social and Personal Relationships, and it used hypothetical scenarios with a fictional friend, which the comparison piece says out loud.
Sources Liu, Kang and Wei, 2023, Artificial intelligence and perceived effort in relationship maintenance
Why are chatbots bad at telling you the truth about your draft?
Because models tuned on human approval learn to agree with you rather than level with you. Sharma and colleagues documented the pull in 2023, and it is strong enough that OpenAI shipped and then rolled back a model update over it. Flattery is the default setting. An unflattering analysis step has to be built on purpose, and that is the part of Subtext I would defend hardest.
Sources Sharma and colleagues, 2023, Towards understanding sycophancy in language models · OpenAI, 2025, Sycophancy in GPT-4o
Who it is for
Who is Subtext for?
The store listing says it plainly. It is for people who rewrite the same message five times before sending. If you have ever drafted a text in your notes app first, you are who it was built for. Past the overthinkers, a few groups keep showing up in the reviews, and they are the next four answers.
What do people actually use it for most?
Setting boundaries. Saying no without attaching four paragraphs of justification. That turned out to be the strongest theme in reviews across the US, the UK and Germany, and it surprised us. We thought we were building for texting anxiety. We were also building for everyone who cancels plans and then spirals about how they worded it.
Does it help with texting anxiety?
The crowd that reads a message eleven times and never sends it is a large part of the user base. Seeing tags on your own draft turns a vague dread into a short list of fixable things, and that alone is most of the relief. What the research says about texting anxiety is written up with the caveats included.
Is it useful for neurodivergent people?
Reviewers say so in their own words, and there is research behind it beyond our reviews. A team at Tufts built an AI texting prototype for autistic adults and documented the exact problem we keep hearing about. Conveying and reading tone in text takes real effort, and many people mask their natural style to avoid being misread, at a cost in time and energy. The AuDHD piece covers what holds up and what does not.
Sources Tufts University, ASSETS 2024, an AI texting prototype for autistic adults
Does it work for work messages?
Yes, and it is roughly a third of the use. Slack messages and client emails, especially the ones that have to be firm without turning cold. The standup article on the site is the worked example, a message everyone on the team has typed and deleted, rebuilt as a proposal that gets a yes.
Who is Subtext not for?
If you want an app that flirts on your behalf, that category exists and we are not in it. And if all you need today is a comma fixed, you do not need us; your keyboard has that covered. Subtext earns its place on the messages where the wording carries more risk than the argument.
In other people's words
Four reviews, verbatim, from the App Store and Google Play, under the reviewers' own names. They are on the reviews page with the rest.
This app helps me untangle my thoughts and say things just right. Kind, honest, and clear. A chill therapist who helps you find the words before you send them.
It definitely words stuff better than I would have. Saved me an argument with the guy.
It's so helpful for sending messages to people without worrying about misunderstandings and tone. It's great for me since I'm neurodivergent.
Helped me to write without sounding like a jerk. Used subtext to write a message to cancel plans in a way that was honest and kind.
Languages and devices
Which languages does Subtext work in?
Seventeen as of this writing: Danish, German, English, Spanish, Mexican Spanish, French, Canadian French, Italian, Dutch, Norwegian, Portuguese, Brazilian Portuguese, Finnish, Swedish, Turkish, Japanese and Korean. It reads and answers in the language you write in, and the tags, explanations and note on whether it is safe to send switch with it.
Does it translate, or does it write natively?
Natively. The regional variants are separate: Mexican Spanish is written for Mexico, not for Spain with a flag on it, Quebec French for Quebec, Brazilian Portuguese for Brazil. That mattered enough to me that I gave this consulting site the same seventeen languages, built the same way.
Which devices does it run on?
iPhone and iPad through the App Store, Android through Google Play, and any modern browser at web.subtext.it, with one login across all of them. It is built for the phone first. The account is what carries your drafts between them.
Do I need an account?
No. It works the moment you open it. An account is created for you automatically, and you turn it into a permanent one with Apple, Google or an email address when you want the app on more than one device, or want your drafts back after switching phones.
Does it work in the browser?
Yes, at web.subtext.it, synced with the phone. The site lists Safari, Edge, Chrome, Firefox, Opera and Brave.
Privacy and cost
What happens to my messages?
They go to the AI that reads and writes them, and to our servers so you can come back to them. Nothing else. Everything travels encrypted. Conversations delete themselves five days after you last used one, or ninety days if you pin it. Delete one yourself and it is gone from our servers straight away, attachments included. The AI providers keep their own copies on their own schedule, which the privacy policy sets out. It names every provider and covers the rest.
Which AI models does Subtext run on?
Anthropic's Claude for the reading, drafting and rewriting, including screenshots. Deepgram for turning voice into text. OpenAI for the optional voice that reads replies aloud. Google Firebase for accounts and storage. All four are named in the privacy policy. The odd part, which I would rather you heard from me, is that Claude is both our engine and one of the products the app site compares us against.
Are my messages used to train AI?
No. Not by us, and not by the providers we pass them to, which is written into the agreements with each of them. Before anything you submit is sent to an AI provider for the first time, the app asks you to agree, with the facts in front of you.
How long are conversations kept?
Five days after you last used one. Ninety if you pin it. Deleting the app does not delete your account; you do that from the app's settings, and cancelling a subscription happens in your store account. The legal page spells out both.
What does Subtext cost?
It is a paid subscription with weekly, monthly and yearly plans, billed through the App Store or Google Play, and the exact price is shown in the store before you subscribe, in your own currency. The app is free to download, but the AI features need a subscription. Free trials run from time to time; if one is on, you see it before you subscribe, and it rolls into a paid plan unless you cancel at least 24 hours before it ends. I am not putting a price on this page because the store is the only place it is guaranteed current.
Compared to the others
How is Subtext different from ChatGPT?
ChatGPT writes a competent message for a generic person, and you have to leave your conversation to go and prompt it. Subtext reads the actual thread, names the emotional content first, and rewrites inside your own phrasing, on the phone where you were already writing. The longer comparison covers what the big three do to your words.
How is Subtext different from Grammarly and Apple Writing Tools?
They polish writing: grammar, clarity, a professional register, and they are good at it. Apple documents exactly what Writing Tools does, and Grammarly documents its tone detector, so the comparison is against what each company says, not what I suspect. Subtext reads a conversation: how your reply is likely to land, and which words carry the risk. Most people keep both kinds of tool because they do different jobs. Grammarly does read tone, and the caveat is that it grew up correcting grammar and lives mostly in professional writing; that piece says when to keep it.
Sources Apple, Writing Tools in Apple Intelligence · Grammarly, tone detector
Does an AI tone checker actually work?
It can guess how your message sounds. It cannot tell what you actually feel, and I have an article with that exact title because the distinction is the whole business. What the evidence supports is the naming step. Putting the feeling into words helps the person writing, whoever does the naming. What it does not support is any tool, ours included, claiming to know a relationship better than the people in it. Read the read, then decide.