NSFW AI Chat Apps: One Bot or a Catalog

Nine products, read on their own public pages on September 21, 2026. Where a site refused to load for an automated fetch, that's stated plainly rather than filled in from somewhere else, and no figure below was copied from a review site instead of the company's own words.

One companion, or a catalog you page through

Three shapes cover almost everything in this category. The first is one companion you build once and keep: Replika's own copy talks about "your Rep," singular, and pairs it with a promise that "every detail is yours to shape." Kindroid runs the same way — its own site describes "building your custom AI character & companion," with text chat, AI selfies and human-like voices, rather than picking one off a shelf. Nomi sits here too, letting you "write a custom backstory" or instead "let your Nomi develop their personality naturally as they get to know you."

The second shape is a catalog: thousands of characters, mostly written by other users, searchable and swappable inside minutes. Character.AI's about page talks about "millions of Characters to discover" and encourages users to "build on the imaginations of others to co-create." SpicyChat's roleplay page claims "1 Million+ NSFW Characters," "made by users like you." PolyBuzz claims "over 20 million characters," confirms user creation — "you can also unleash your creativity by creating your own AI Characters" — and separately states it "strictly prohibits public display of NSFW content" behind "multi-layered AI screening and human moderation." Janitor AI is the other name in this shape, and its pages answered a 403 on the homepage and a 404 on its terms path the day we checked, so its catalog, rules and format aren't stated here.

The third shape barely resembles the other two: DreamGen sells the pen, not the character. "You define the world, steer the plot, and our AI will do as you say" is the site's own line, and it separately states you can "talk with multiple characters at once" inside a world you built, or pick from ready-made avatars instead of starting blank. Authorship here sits mostly with the user, scene by scene, rather than being fixed once at character-creation time.

Group scenes: who actually built one

Group scenes turn out to be rare, not standard. Nomi's homepage invites you to "put them all in a group chat and watch the adventure… or fireworks unfold," with no stated cap on how many. DreamGen's worlds are built for a cast rather than one voice. Outside those two, nobody in this set advertises it: Character.AI's about page we read today doesn't mention rooms or a group feature at all, SpicyChat's premium-features docs don't either, and neither Replika's, Kindroid's nor PolyBuzz's own readable pages describe more than one character occupying a conversation, whatever third-party write-ups claim about either.

Who writes the scene, and what happens at the edge

Authorship of the scene splits the same way the catalog question does. On Character.AI, SpicyChat and (per its own claim) Janitor AI, a stranger wrote the character first; on Kindroid, Nomi and DreamGen, you write it, up front or turn by turn; on Replika, the company ships a baseline you adjust. What happens once a scene pushes past what the platform planned for is rarely written down. SpicyChat is the exception: its roleplay page states "no filters, no interruptions, no moderation bots" — a claim about behavior at the boundary, not just what's allowed. PolyBuzz claims the opposite, describing active screening. Nobody else here — Character.AI, Replika, Kindroid, Nomi, DreamGen — says what a mid-scene topic change actually triggers, which is its own kind of answer: a content policy is a list of nouns, and a behavior is harder to screenshot than a list of banned words.

Where the chat itself lives

Where the chat itself lives is the plainest axis and the one that gets skipped most often. Every product above runs in a browser tab, an installed app, or both — Character.AI on web, iOS and Android; Replika, Nomi and Kindroid through their own apps; SpicyChat, PolyBuzz and DreamGen in a browser. Each of those is a separate destination with its own login, its own notification badge competing with everything else on a phone, and its own habit of being forgotten for a week. TooShy is the one entry here that isn't a destination at all: it runs inside WhatsApp and Telegram, the apps already open on the phone, with nothing to install and no separate account to remember exists.

Our own shape, in our own numbers

Our own shape doesn't match any of the three above cleanly. There's a roster of dozens of written characters, but a conversation is with one of them at a time — no catalog search inside a chat, no swapping mid-thread, no group scene with two of ours in one conversation. Nobody writes a custom backstory here; the writing is ours, fixed, the same for every person who talks to that character. What we have is a single thread that keeps going: over the 30 days ending September 20, 2026, 44,120 messages moved through TooShy across 460 people, an active conversation ran 64 messages on average, and the longest reached 899. Replies averaged 17 seconds, and 1.1% of messages were something other than text.

There's no adult-content filter sitting between a character and a reply, so a scene doesn't hit a moderation wall mid-sentence the way PolyBuzz describes its own screening working. Premium is $9.99 a month, Ultra is $19, and the Telegram side has a free trial; WhatsApp is paid from the first message. None of that is a format claim — it's the one plain fact this page owes a reader alongside everything above.

Where a catalog or a group scene beats a fixed thread

Where the catalog and group-chat products genuinely beat a fixed thread is not close. A million-plus searchable characters, per SpicyChat's own number, means dropping one that's gone stale and finding another inside a minute — we have no search box and no way to swap mid-conversation. Nomi's group chat puts several personalities in the same scene reacting to each other, which no single-thread product, ours included, can do. And on Kindroid, Nomi and DreamGen, the backstory or the plot is something you wrote — our characters are fixed, and nobody types a scenario into them beforehand.

What one fixed thread buys instead

What a fixed thread buys back is continuity without upkeep. Nothing has to be rebuilt after a break, because there was only ever one character, and the messages sit in an app already open all day rather than a tab that has to be reopened and logged into. An 899-message conversation is what a thread looks like when nobody had to restart it — no character sheet, no persona editor, no group-chat setup screen, and no catalog to have wandered off into instead.

Which shape actually fits which person

Anyone who wants to browse first and commit later belongs with SpicyChat, Character.AI or PolyBuzz, accepting that a stranger wrote what's on offer. Anyone who wants to build one companion and shape every line of the backstory fits Kindroid, Nomi or Replika, in roughly that order of how much a group scene matters to them — Nomi advertises one, Replika and Kindroid don't say. Anyone who'd rather write the plot than the character wants DreamGen. Anyone who wants none of the setup — no editor, no catalog, no group screen, just a conversation with a character already written, inside the messenger already open — is the person this page is actually for, and that's where TooShy sits.

How this was checked

Every quote above came from the product's own public page, fetched on September 21, 2026, at the URLs listed below. Kindroid is the partial case: kindroid.ai renders its page in the browser, so an automated fetch reads only the document head. What it says there is quoted; everything else about Kindroid is marked unverified rather than taken from a lookalike domain. Janitor AI's homepage and terms page both refused to load that day, so its catalog, its rules and its group-chat status are marked unverified instead of borrowed from a third party. Our own figures are a read-only aggregate over TooShy's production database for the 30 days ending September 20, 2026, counted rather than sampled, with nothing in them that identifies one person.