Updated: 2026-07-24
A clone from chat is built from a Telegram chat export: you export the history, drop the files onto the clone page, and the agent starts speaking in the chosen person’s voice. The key privacy point: the export is parsed right in your browser, and only a compact style profile reaches the server — not your messages.
Open the chat in Telegram Desktop. History export is available only in the desktop Telegram app (not mobile, not web). Open the direct or group chat you need.
Start the history export. Click the menu in the top-right corner of the chat and choose "Export chat history".
Choose a format and turn off media. Pick JSON or HTML. Turn off exporting photos, video and voice — a clone needs only text, so the export runs faster.
Drop the files onto the clone page. Open clone creation and drag the resulting folder or the .zip archive. Parsing runs locally in the browser; progress is visible right away.
Pick a participant and build the card. Choose whose voice you need — the platform assembles the character card with a single call to the brain. Review and edit the name, personality, rules and greeting.
Export files are parsed locally, in your browser: the archive is unpacked on your machine, HTML and result.json are read there too. A script deterministically computes a compact style profile — reply length, word and emoji frequencies, frequent phrasings, a small sample of verbatim lines, and the names and facts mentioned in the chat. Only this profile (tens of kilobytes) is sent to the server.
The chat history itself is never uploaded to Corpis servers. This is done not only for privacy: the frequencies are counted from your data rather than invented by the model, and one call to the brain over the finished profile is enough instead of dozens of calls over chunks of history.
After the clone is created its profile is stored on the agent, so the original files are no longer needed for further work. The "Clone source" tab offers two actions. Rebuild — regenerate the character card from the already saved profile, without re-uploading the export (useful when the cloning pipeline has been improved).
Top-up — add another chat (a different conversation with the same person): the new data is deterministically merged with the old (frequencies add up, statistics are recomputed), and the clone is rebuilt on the combined corpus. The more live text, the truer the voice.
Voice fidelity grows with the amount of chat: from a short conversation the profile is not enough to catch the manner. Add more history via top-up. If the clone was created long ago, before pipeline improvements, rebuild the card on the "Clone source" tab: the old clone keeps working, but the new one is assembled by more accurate rules.
One honest limit remains: on direct questions about rare specifics (everyday details, someone’s lore) the clone may improvise — that is the ceiling of the language model itself, not of the profile. The profile sets the manner of speech, not a full biography of the person.
No. 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 never uploaded to Corpis servers.
JSON, HTML and a .zip archive all work; JSON is parsed most accurately. You do not need to export media — only text matters for a clone, and without photos and video the export is faster and smaller.
Yes, the "Clone source" tab has top-up: upload another chat with the same person and the new data merges deterministically with the old, then the clone rebuilds on the combined corpus. The same tab has rebuild from the already saved profile, with no re-upload.