The AI companion that shows its work

Freedom is the feature

Most "AI companion" bots are a script with a picture on top: one persona, one tone, a handful of canned moods, a paywall between you and the third message, and an ad to watch or a channel to join before the paywall. Talaforge is built on the opposite premise — that the interesting thing is the character you write, and that every knob a serious roleplayer expects should be in your hands.

A Talaforge character is text you own. The system prompt is yours to write from scratch or to have drafted for you from a single sentence — name, personality, voice, opening scene — and then edit in every part. There is no single locked "house" persona. You can keep a character private so that only you can ever open it, or publish it so that others find it in the catalogue and through your link.

The same goes for the machine underneath. The seven sampling controls that a local front-end exposes — temperature, top-p, top-k, min-p and the three penalties — are exposed here too, per character and per model, with a slider in the chat, a per-key «send» switch, and a JSON preset you can export and import. If you have a preset from another front-end, it imports as it is. Community-standard character cards import too, as JSON or PNG. On top of the character's own prompt you keep personal prompts at three scopes — global, per character, per topic — that apply only to your chats.

Try it: the sampling lab

This is the kind of control that a scripted bot cannot offer, because a script has nothing to sample. Move the two sliders and watch the same sentence loosen or tighten. The JSON underneath is the exact shape the bot exports and imports.

None of this is hidden behind a premium tier. The editor is free, character creation is unlimited, and a private character costs nothing to keep.

A script has one answer. A character has a temperature.The difference between a chatbot and a character

We name our models

In 2026, the loudest complaint in this category is not "the filter" — it is "the model got dumber overnight and nobody told me." Platforms retire engines, rename them, and swap them under a pet name. Talaforge does the opposite: the model answering you is named on the screen, and changing it is your decision.

Five text models are in rotation, and each one is called by its real name. Standard is the quick, steady default every new character starts on. GLM is the step up for tighter reasoning across a long exchange. Claude Opus carries slow-burn prose and subtext. Google Gemma is fast banter. DeepSeek holds a long, twisting plot together. You switch between them in the chat, mid-scene, with nothing lost: persona, memory and history stay exactly where they were, and only the voice shifts.

Images work the same way. The picker lists the image engines by name — nine at the time of writing, among them Z-Image, Flux, Krea, Seedream and Nano Banana — each with its energy weight shown beside it, so the choice between a fast sketch and a slow, detailed render is yours and visible before you make it. And if you would rather run your own key, you can: a bring-your-own-key card sits beside the five as a sixth choice.

Voice is a third choice, not an afterthought: every character has a voice of its own that you can audition before you pick it, and replies can be silent, spoken on request, or spoken automatically.

When a model degrades on the provider side, a router fails over silently to keep the scene going — and it tells you what it did. The point is not that we never have a bad day. The point is that you always know what is answering.

Try it: the picker

The keyboard below mirrors the one in the bot. Tap a model to hear the same beat in a different voice; switch to images and voice to see the other two choices.

You are not buying a cat in a bag. The name of the model is on the chip, and the chip is a button.Why the model picker is a promise, not a menu

Built for Telegram, not ported to it

A lot of bots are a web app squeezed through a chat window. Talaforge is written for the things Telegram actually does well in 2026 — streaming drafts, forum topics, rich text, native voice — and it uses all of them.

Replies in the main bot stream as they are generated: the message grows word by word, the way a person types, and a Stop button is there while it does. You do not stare at "typing…" for thirty seconds and then receive a wall. If you stop it, you keep what was written.

Formatting is real. Bold, italics, quotes and lists render as Telegram renders them and code stays in code blocks; long replies continue cleanly across messages instead of being cut. In the standalone bots you make (chapter seven) the full rich-text set is on by default — headings, tables, spoilers, footnotes, and formulas that arrive as formulas, not as a row of backslashes — because a character who is a tutor needs it.

Every scene lives in its own forum topic. Open the bot as a forum and each character — or each branch of a story — is a tab with its own name, its own pinned card and its own history, so ten conversations do not collapse into one scroll. Switching a character is switching a tab.

Media runs both ways in the same chat. A character can send a photo of the scene you are in and answer in a voice message; you can send a photo and the character sees it and answers in character about it.

And there is nothing to install. Your Telegram account is the account. A shared character opens as a chat, not as a download page followed by a signup form followed by an age scan.

That last point is bigger than convenience. A standalone app answers to its publisher's store before it answers to you: the store's content rules decide what a character may say, a review queue decides when a fix ships, a 2026 policy wave against "wrapper" AI apps decides whether it stays listed at all, and the store's cut is priced into every credit you buy. When a store changes its mind, whole companion apps have been retired with a month's notice and their users left holding an empty icon. A Telegram bot has none of that layer: the character you are talking to today is not one policy review away from disappearing, and nothing about how you pay carries a store's margin.

The demonstration at the top of this page is what streaming looks like from the reader's side: the reply appears as it is written, the model's name sits on the chip, and Stop and Regenerate are one tap away. What an AI roleplay bot for Telegram actually is covers the three ways to get one and why an already-running one is usually the right answer.

If it is a Telegram bot, it should feel like Telegram. Topics, drafts, voice notes, rich text — all of it, not a subset.The house rule for the chat surface

Fork the scene. Keep both.

The single most useful thing a story tool can do is let you ask "what if she had said no?" without destroying the version where she said yes. Talaforge is the only Telegram bot we know of that answers this with real branches: fork any reply into a new forum topic, and both threads stay alive.

Under every reply there is Regenerate, Edit and Fork. Regenerate re-rolls the same beat. Edit rewrites the character's line — or your own — and the story continues from the rewritten text. Fork does the interesting thing: it opens a new topic with the whole history up to that point copied in, and continues from there. The original tab is untouched. You can fork a fork.

Memory follows the branch: facts the character learned before the fork are inherited by both threads, and facts learned afterwards belong to the thread where they happened. Nothing bleeds between branches, and nothing is lost when you decide the other one was better after all.

Try it: fork the lighthouse

Press Fork under the last line. A second tab appears with the same history and a different continuation. Switch between them — both are live.

That is the whole mechanic, and it is why writers use Talaforge to test dialogue: two versions, side by side, in the same chat, with the same character. How memory is inherited when a scene branches goes into the details.

Regenerate asks the same question again. Fork asks a different one — and keeps the first answer.Three buttons under every reply

Memory that condenses and retrieves

Every "the character forgot me" complaint resolves to one of two things: the context window filled up, or nothing was ever stored. Talaforge handles both — older turns are condensed into episodes before they fall out of the window, and stored facts are retrieved back into the scene when they matter. Ultra-long conversations are the normal case, not the edge case.

Layer one is condensation. When a scene outgrows what the model can hold, the oldest messages are not simply dropped: a one-shot pass turns the window that is about to fall off into a short episode — a paragraph and a few grounded facts — and that episode stays in the character's memory. The condenser is forbidden from inventing; a bullet that does not appear in the original text is thrown away rather than kept.

Layer two is facts. As you talk, the character extracts short, readable lines — a name, a preference, a thing that happened — and stores them per character, per topic, per person. Emails, phone numbers and card numbers are stripped before anything is written. When a later message touches one of those facts, it is retrieved and put back in front of the model.

Layer three is you. Every stored fact is a line you can open, correct or delete from inside the chat. Correcting a memory line does what people hope "training" would do, instantly and visibly.

Try it: condense a scene

The transcript on the left is what the model sees. Press Condense and watch the oldest turns fold into an episode while the facts stay put; then ask a question and see which facts light up.

There is a longer, plainer explanation of what is stored and where in Memory you can open, correct and delete, and a field note on why you don't train an AI character — you correct its memory instead.

The window forgets. The memory doesn't — and you can read every line of it.Condensation, retrieval, and the delete button

Photos with a locked look, a real voice, and eyes

The three media capabilities are shipped, on by default, and priced in energy you can see. The character sends photos of the scene you are actually in, with the same face every time; answers in native voice messages; and sees a photo you send and replies in character about it.

The look is locked to the character, not re-rolled per picture. That is the difference between an illustration and a companion: ask for a picture in a café, then in daylight, then in a period drawing room, and the same person appears in all three. Nobody uploads a reference each time; the reference is the character. Six real frames from one conversation are below — the same six the field notes carry.

The engine is a choice, per image, by name, and the energy weight is shown before you press. A render that fails or is blocked is not charged, and a daily ceiling protects the balance.

Voice replies are Telegram voice messages — waveform, playback speed, forwardable — not a widget. Each character has a voice of its own; you audition it before you choose. Off, on request, or automatic, per character or globally.

And the character can look: send a photo into the chat and it is described back to you in character. That photo is relayed to a third-party model to be understood, and the data-processing policy says so in plain words.

The full account — what needs no configuration, what is billed how, and what is deliberately not claimed — is in An AI character that sends photos and speaks, with the six frames at full size in One character, every scene.

Same face in every scene. That is not a filter setting; it is what a character is.The consistency rule for generated images

Your character, its own bot

Any character can become a separate Telegram bot with its own @username in a couple of minutes — without code and without @BotFather. The bot is created from inside Talaforge through Telegram's managed-bot mechanism; the only screen outside Talaforge is the one where you pick the @username. Telegram creates it, so it sits in your account like any other bot; Talaforge runs the character behind it. This is where the platform stops being an app and becomes a family of bots.

This runs on the newest part of the Telegram platform — bots that create bots — and Talaforge was built on it rather than on the old ritual. The old way, which is still what most tools ask of you: open @BotFather, run a command, copy a token, paste the token into a chat. Here there is none of that. You tap Create my bot, Telegram opens its own creation sheet with a suggested @username, you keep it or type another, Telegram tells you on the spot whether it is free — and the new bot is handed straight to Talaforge. The key never travels through a message: Talaforge fetches it from Telegram directly and can rotate it later without you.

Then Talaforge dresses the bot as the character, automatically: the name, the profile picture from the character's own avatar, the description on the «What can this bot do?» screen, a menu button — and the first message a stranger receives on /start is the character's own greeting, sent without spending a single unit of energy. A bot that opens with an empty screen is what the platform gives you by default; a bot that opens as your character is what you get here.

It inherits the character: its persona, its memory, the model behind it, the rich formatting — and the character stays editable and swappable afterwards, so the bot never becomes a frozen copy. It can be open to anyone who finds the handle, or only you — a private line to your own character with an address you can reach from any device.

Everything about it is run from one place — Profile → 🤖 My bots inside Talaforge: stop and resume, swap the character, open it to everyone or close it to everyone but you, allow groups, let it write the first comment under channel posts, let it join a conversation on its own now and then, switch photos and voice on, set the adult-photo switch, set your cut of what people top up inside it, check the connection and reissue the key, reset the look to the character, top up its energy. The one thing that stays with Telegram is deleting the bot for good and freeing the @username — that is Telegram's to do, and it cannot be undone.

It can be added to a group and host the conversation in character, answering the people who address it rather than spamming the room. It can connect through Telegram Business and answer your own incoming messages in the character's voice while you are away. It can open and answer the comment thread under a channel post.

The worked example is Tavi, a tutor character who sits in a study group and answers when a student addresses him — checking a solution step by step, finding the wrong line, explaining it in the group's own language, and staying quiet otherwise. Tavi joins the group chat shows a real session, and the piece is written separately for each market rather than translated once.

And it earns. People who arrive through your link are attributed to you; a share of what they spend comes back in Telegram Stars through Telegram's own affiliate programme, and Talaforge adds an energy bonus of its own when a referred person makes their first purchase — on top, never instead.

Try it: from a character to a live bot

Four steps. Tap any of them; the sequence plays on its own — and watch what Talaforge fills in for you.

The honest version of the revenue story — who earns meaningfully and who does not — is in Earning from a character you wrote, and the mechanics in Your character, its own Telegram bot.

It is not a marketplace. It is a bot with your name on it, and a share of what happens inside.What "your own bot" means, precisely

The App: a catalogue, a cabinet, and your links

Next to the bot there is a Mini App — the same account, opened as a screen. It carries the catalogue of characters and standalone bots, a personal cabinet with every charge and top-up itemised, settings and payment, and the referral tools.

The catalogue is multilingual by construction: characters carry the language they were written in, and the App shows you yours first. Each listing has a rating from the people who actually talked to it — never shown without its vote count — a heart count, and a row of similar characters worked out from tags that describe what a character is rather than from what is popular. Standalone bots appear alongside in-bot characters, so a character that has grown into its own @username is one tap away either way.

The account is the boring part done properly. The cabinet shows the balance, your characters, your bots and your invite link; the bot's transaction screen keeps every energy charge as a line — which character, which model, which image engine, how much — and so every top-up, every free-energy claim and every referral bonus. No mystery deductions, no surprise purchases — and no line that says «earned by watching an ad», because that line does not exist here.

Links are referral links. Every character link you share attributes the person who opens it to you. Telegram's own Stars affiliate programme does the accounting and the payout; Talaforge does not touch it or bend it. What we add is ours to add: an energy bonus to the person who brought someone who then made their first purchase. The only thing an author has to remember is to paste their own affiliate `tgr` token into their creator dashboard once, so every share link wraps itself.

Try it: the catalogue

Tap a character to see who is similar. The recommendation follows the tags, not the hearts.

Every link is a referral link, and the accounting is Telegram's. We only add on top.The rule for the creator economy

Pay the way you already pay

Two things about paying that almost nobody else in this category does at once: three routes — Telegram Stars natively, bank card and crypto through a partner — and one honest model — every pack is an energy pack, spent only as you use it, and every top-up grants a month of Premium on top. There is no subscription to keep alive and no tier to pick.

Stars are the native route: bought inside Telegram with whatever your phone already uses, spent on energy in one tap, and never showing the bot a card number. If your phone's store is not an option, a partner sells Stars for a bank card, a local bank transfer, or crypto, and the button for it sits right next to every energy pack. It is a full set, and it is rare — some services take only cards, one or two take only crypto, most take only their own app store.

There is one thing to buy: energy. Every pack is an energy pack, and energy is spent only on what you actually use — a message on a heavier model, an image, seconds of voice — with a failed render never charged. Nothing expires unused and nothing renews behind your back. Premium is not a second product: any top-up grants it for thirty days, and it unlocks the premium characters and the doubled free energy. It is a thank-you for topping up, not a plan you have to choose.

Free energy is real and scheduled. Everyone can claim it weekly. Anyone on Premium — that is, anyone who has topped up in the last thirty days — receives it automatically, at double the rate, every three days. A daily ceiling protects everyone from a runaway evening.

And it is claimed, not earned. There is no "watch an ad to continue", no "subscribe to these five channels to unlock", no task wall, no daily quest, no mid-chat advertisement — the things a lot of free bots in this category run on. Your attention is not the currency here; energy is, and it is either on the schedule or in your balance.

Try it: the three routes

Flip a card to see what each route actually does. The switch underneath is not a toggle — both sides are on, and that is the point.

What "free" covers — unlimited character creation, no premium editor, no card needed to start — is spelled out in Create your own AI character, free.

Stars, card, crypto. One balance, spent as you use it — and a month of Premium with every top-up.The payment rule

36 languages, one year, no drama

Talaforge speaks 36 languages in the bot and the App, auto-detected, right-to-left native — Arabic, Persian, Hebrew and Urdu are first-class, not fallbacks. It has run in production since autumn 2025 with thousands of monthly users, no data incident and no scandal, which in this category in 2026 is itself a feature.

Language is not a translated menu. Characters are written in the language of their author, the catalogue shows you yours first, and the bot answers in the language you write. This site exists in 35 of the 36; the field notes are published in all of them at once.

On privacy, the posture is plain rather than loud. Messages are processed by AI models to generate replies; what is stored is chat history plus per-character memory that you can open and delete line by line; payments run through Telegram and a partner, so the bot never sees a card number. The whole policy is one document, in every language, at one URL, and it changes in the same pass as any feature that touches your data.

On reliability, the router fails over between models when a provider has a bad hour, the balance has a daily ceiling, and a failed render is not charged. A year of that is why the timeline further down is mostly features rather than apologies.

If you are choosing between platforms, AI character platforms, compared honestly maps the space by shape without naming winners, and Why this exists tells the origin story.

Thirty-six languages, one policy, one balance, one account — the one you already have.The shape of the thing

What a high-end Telegram AI bot needs — and who has it

No competitor is named here, deliberately: shapes age better than brands, and the point is a checklist you can carry to any platform. The right-hand columns describe the typical bot and the typical web app in this category in 2026.

The table is a checklist, not a verdict. If a single long-term companion is what you want, a companion app may still serve you better — the honest comparison says so. If you want to direct characters, pick the model, edit the memory and keep every branch, this is the shape built for that.

How it got here

Month by month, in the order the features actually shipped. The dates are the product's, not a roadmap's.

Frequently asked

Short answers first, the long ones are linked.