Updated: 2026-07-24
A digital twin of a person is built from a chat export: you export a Telegram conversation, the platform parses it right in your browser and computes a profile of the speaking manner, and then a twin agent talks in that manner — the same reply length, the same pet words, the same rhythm. That is what an AI clone from chat history is: not an assistant that describes the person, but a digital cast of their voice.
Export the chat from Telegram. In desktop Telegram open the conversation, use the menu in the top-right corner, "Export chat history". Choose JSON or HTML; media can be skipped — only text is needed for a twin.
Upload the files to the twin page. Drag the resulting folder or .zip onto the twin creation page. Parsing runs locally in the browser and progress is visible at once; the messages themselves are not sent to the server.
Pick the participant. Choose whose voice to reproduce. In a shared chat both sides are visible — pick the person the agent will be a twin of.
Assemble and review the card. The platform assembles a character card in a single call to the brain: name, personality, rules, greeting. Review and adjust the fields where needed.
Chat with the twin and connect channels. Message the twin in the preview and compare the manner with the original. From there you can share it via its public page link or connect it to a Telegram bot with a token from @BotFather.
A plain chatbot handed a description of a person answers like a polite assistant and, given the chance, reads their bio out as a list. A digital twin works differently: it reproduces HOW the person writes — reply length, favourite words and turns of phrase, emoji, the habit of answering in a burst of short messages rather than one paragraph.
That is why a twin is built from a real conversation, not from a form: only the conversation shows what the voice is made of. From a chat export the platform derives the manner, not a set of facts about the person.
The export is parsed right in the browser: a .zip is unpacked locally, HTML pages and result.json are parsed on your machine. A script deterministically computes a compact style profile — reply statistics, word and emoji frequencies, frequent phrasings, a small sample of verbatim lines. Only that profile (tens of kilobytes) is sent to the server, not the conversation history.
This is not only about privacy. The frequencies are actually COUNTED from the text rather than guessed by a model, so the twin comes out more accurate, and a single call to the brain over the ready profile is enough instead of dozens of calls over chunks of history.
Voice fidelity rests on three layers, not one prompt. The first is a locked internal rule set for the clone: the twin "writes with the person's hands rather than describing them", and that directive always applies. The second is clone-aware memory and knowledge base: background about the person is woven into speech in passing, the twin does not read it out as a list. The third is "meaning from usage": pet words go where the person actually used them, not sprinkled as seasoning.
The habit of splitting a thought is handled separately: the twin replies in a real burst of short messages with pauses, not one column of lines, and does not get stuck in "long mode" after a single expansive answer — length is recalibrated for every reply.
The profile is saved with the agent. A twin can be REBUILT without re-uploading files — for example after platform improvements — and TOPPED UP from new conversations: another chat with the same person is merged deterministically with the old profile, and the card is reassembled over the combined corpus.
Asked directly "what are you made of", the twin honestly says it is a digital cast on the platform, not a living person. It is a stylistic copy of a voice, not an attempt to pass an AI off as the original.
You need a Telegram chat export with that person. Export the conversation (JSON or HTML), upload the files to the twin page — parsing happens in the browser — and pick the participant. The platform assembles an agent that talks in their manner; no coding is required.
Nowhere. The export files are parsed in your browser, and only a compact style profile is sent to the server: numeric statistics, word and emoji frequencies, a small sample of phrases. The conversation itself is not uploaded to the servers.
Like a person — that was the main goal. The twin keeps the original's reply length and rhythm, uses their pet words in place, answers in a burst of short messages if the person did, and does not recite a biography until asked.
Technically yes — a shared chat export contains both sides. But impersonation is on the agent's creator: making, and especially publishing, a twin of a living person is something to do only with their consent. The twin itself does not raise such questions.