Beta · Last updated 2026-11-16

Privacy Policy

This policy explains what information is collected when you use the Haldun AI home page and studio, how it is used, who it is shared with, and what rights you have. Haldun Ai LLC is responsible for the information. We follow the Law of Mongolia on Personal Data Protection.

This is an English translation for convenience. If it differs from the Mongolian version, the Mongolian version prevails.

In short

1. What information, and where

1.1Account. Email address, sign‑up and last sign‑in dates, credit balance and credit history (date, amount, reason, creation number). Each time you sign in, a 6‑digit code is sent to your email. We hold no password. If you sign in with Facebook or Google, that service passes us (through Supabase) your email, name and profile picture link; we use only the email for your account. We never see your Facebook or Google password and never post anything in your name. You can remove the connection at any time in Facebook’s “Apps and websites” or Google’s “Third‑party connections” settings; to have your account deleted, see section 7. If you buy a plan or extra credits: the plan (or the number of extra credits), its period (month or year), amount, date, QPay invoice and transaction numbers, and when your plan ends (used to close credit spending after a plan ends, until you renew); for a yearly plan, the schedule of its monthly credits (how many, when, whether given, and how many were added under the 3‑month balance cap).

1.2In your browser (never sent to us). Studio settings, creation history, character pictures and a guest’s (not signed in) “My library” are kept in the browser’s storage (localStorage, IndexedDB). The word fixes you make with ✏ in Quality translation (wrong → right, up to 200) are kept only in this browser and your browser applies them to your next subtitles — they are never sent to us; corrected subtitle lines are sent only when you press 🔥 to burn them into a video again (Modal in the table below). Clearing the browser’s site data deletes them. When you sign out of your account, this device’s creation history, the ideas you wrote, local copies of creations, the AI editor project, characters, boards, remembered subtitle fixes and your email address are deleted (display settings stay).

1.2a“My library” (signed in). The pictures and videos you create, the idea you wrote for them (up to 300 characters), tool, aspect ratio and date are stored in our database and file storage (Supabase), linked to your account — so you see them from any device. Only you can view the files, through links that expire after 1 hour; they are not public. When a “Quality translation” job starts, an encrypted job note (file name, tool) is kept in your folder for a short time — so even if you reload or close the page, the result reaches “My library” the next time you open the site; it is deleted when the job ends; otherwise it expires after 3 hours and is deleted the next time you open the site.

1.2bThe Showcase. If you publish a template: your public name, a copy of the work (in a publicly readable file store), its cover, title, description, category, recipe (text and settings), date, and its heart, bought and use counts. We also keep your hearts, comments, reports and the templates you got (when, and whether you had bought credits at that time). Authors do not see who got their templates — only how many.

1.2cThe AI editor. The video you add and your edit (timeline, text) are kept only in your browser (IndexedDB) and restored when you reopen the page — they do not reach us. When you press “Cut with AI”, the browser takes only the audio from the recording (16 kHz WAV) and uploads it to our file store for a short time. A GPU service turns the speech into text with a time for every word, and AI finds the parts to remove in that text (fillers, stutters, repeated takes, corrections). The audio file is deleted when the job ends; the text and cuts go back only to your browser. You decide whether to save the finished video to “My library” (1.2a).

1.2dMake 4K. When you make a video from “My library” or a finished AI editor video 4K, that video goes to a GPU service, is upscaled with AI and saved as a new file in your “My library” (1.2a). The temporary copy sent from the editor is deleted when the job ends; nothing is kept on the GPU service.

1.2eBackground. When you replace the background in the AI editor, that video (or, if it is in “My library”, a link to it) goes to a GPU service; AI cuts out the main subject (a person, an animal…) in every frame, replaces the background (or, for “✨ HD” on V2, leaves it transparent) and saves the result as a new file in your “My library” (1.2a). The temporary copy sent from your computer is deleted when the job ends; nothing is kept on the GPU service. The original recording stays in your browser.

1.2fThe Haldun guide. The helper in the corner of every page. A question you write in the chat (up to 400 characters) and the name of the page you are on (for example “Tools → Audio”) go to our server and from there to an AI service, which produces the answer. If you choose “Say it”, the answer text goes to ElevenLabs to be turned into speech. On phones there is no chat: when you press and hold the guide and say the name of a button, your browser records up to 3 seconds from the microphone (it stops sooner if you lift your finger), and the recording goes through our server to ElevenLabs to be turned into text. Our server matches that text with the button names and picks a fixed explanation — nothing goes to an AI service. The recording and the heard text are not kept or logged (only which explanation was picked). The microphone is used only while you press and hold the guide. Questions and answers are not kept and not written to logs (only the answer’s length and whether it was spoken) — the chat history disappears when you close the page. Daily limits are counted as described in 1.5. The spoken tour lines are the same for everyone and contain none of your information. Your choice to hide or mute the helper is kept in your browser (1.2).

1.2gRelight. When you relight an image in the AI editor (for a video, its current frame), that image goes to Google’s image AI (Nano Banana Pro) and comes back as a new picture for each lighting style you chose; the pictures are saved in your “My library” (1.2a). Our server does not keep the original — it stays in your browser.

1.2hImage → Edit. When you choose “Restore” or “4K size”, your picture is copied to our file store for a short time and goes to a GPU service; AI restores faces and fine detail (or makes it 4K), upscales it and saves the result as a new picture in your “My library” (1.2a). The temporary copy is deleted when the job ends; nothing is kept on the GPU service. When you choose “Change”, your picture and your request (if you wrote one, translated into English) go to the picture AI you chose (OpenAI’s GPT Image or Google’s Nano Banana Pro), and the edited picture is saved in your “My library”. Our server does not keep the original. If you turn on more than one part, the picture goes first to the picture AI (“Change”) and then to the GPU service (restore, 4K); only the final picture is saved in your “My library”, and the in‑between picture is deleted when the job ends. When you choose “Erase”, your picture and a mask of the part you painted with the brush (a black‑and‑white picture) go to the GPU service, which fills the painted part from its surroundings; if you wrote what should be there instead, your picture, the mask and your text (translated into English) go to OpenAI’s GPT Image, and the GPU service puts only the painted part of the result back into your picture. The unpainted part does not change; the final picture is saved in your “My library”, and the temporary copies are deleted when the job ends.

1.2iAI editor → Product. The 1–3 product pictures you choose and your description (if you wrote one, translated into English) go to ByteDance’s video AI (through BytePlus) and come back as a product video, which is saved in your “My library” (1.2a). Our server does not keep the pictures.

1.2jAI editor → Cover. The title, artist, genre, mood and description you write go from our server to Anthropic’s Claude, which turns them into 4 cover ideas (in English); OpenAI’s GPT Image then draws the 4 covers in one picture from those ideas and the title, artist, genre and mood, and the picture is saved in your “My library” (1.2a). If Claude is unavailable, the description also goes straight to GPT Image. When you pick a cover, your browser cuts it out and it goes to the GPU service to be made 4K (like “4K size”, 1.2h); the 4K cover is saved in your “My library”. If you add the artist’s logo, it does not go to our server: your browser places it on the cover you pick, and the cover with the logo goes to the GPU service to be made 4K and is saved in your “My library”. If you press “✨ Bring to life”, your browser sends the picked cover without the logo (in the middle of a 9:16 frame), with the title, artist, genre, mood and description, through our server to Anthropic’s Claude, which directs what should move; Google’s video AI (Gemini API) then makes a short video from that direction and the cover picture, and the video is saved in your “My library”. The logo does not go to our server here either: your browser places it on the video, makes a 9:16 and a 1:1 version and saves the 9:16 one with the logo in your “My library”. The title is kept as the creation’s name; our server does not keep the rest of what you wrote.

1.2kAI editor → Video story. The topic you write (and the style picture, if you choose one) goes from our server to Anthropic’s Claude, which turns it into a 4-shot script: a Mongolian voice-over, the motion, the look of each shot and a music direction (in English). OpenAI’s GPT Image draws the 4 shots in one picture from those looks (and the style picture, if any), and the picture is saved in your “My library” (1.2a). When you approve, your browser cuts out each shot and sends it with its motion text to Google’s video generation service to become a clip (like video, 1.3); the voice-over lines go to ElevenLabs (like voice) and the music direction to ElevenLabs’ music. If you choose “My own” music, the file itself is not sent to us — it is only mixed into the finished video in your browser. The clips, the music and the finished video your browser puts together are saved in your “My library”; the topic is kept as the creation’s name, and our server does not keep the rest.

1.2lTools → Video → Dance. The action you write for each shot, “Who” and “Where” (if written, translated into English by Anthropic’s Claude), the dialogue (as you wrote it), the camera you chose, the Kam code, your options, the first frame and reference pictures (up to 3) go to ByteDance’s video AI (through BytePlus) and come back as a video, which is saved in your “My library” (1.2a); the first shot’s action becomes the creation’s name. Our server does not keep the pictures. If you add a Haldun Cam recording, it is copied to our file store for a short time and goes to the video AI through a link that expires in 1 hour; it is deleted when the job ends (even if it fails).

1.2mMNNZN Studio — bookings. When you send the booking form on the MNNZN Studio site (haldun.ai/mnnzn/), your name, phone number or email, the service, your note and the page address go from our server through a Telegram bot to MNNZN Studio’s chat — only so the studio can reply to your booking (arrange a time). If you are signed in to Haldun, the site shows your email in the top bar and fills it into the form’s contact field (you can change or delete it) — nothing is sent until you press “Send”. Our server does not keep bookings; to protect against too many submissions it counts IP addresses in memory for 1 hour.

1.2nDance → 🕺 3D characters. “3D from a picture”: the character picture you add (made into a JPEG of up to 2048 px by your browser) goes from our server to Tripo’s (VAST) 3D service and comes back as a 3D model (.glb) with a skeleton plus idle and walk motions. Only add pictures of characters made with AI or drawn by you — photos of real people are not allowed (Terms 3.10). Our server does not keep the picture. “Motion from video”: the video you add (MP4, MOV, up to 30 seconds) is copied to our file store for a short time and goes from our server to Uthana’s motion service, which reads the body motion of the person in it and puts it on the character you chose (the skeleton-free model of your 3D character, or the ready-made grey character) as a .glb. The video is deleted from our store once it has been sent (even if the job fails). The finished .glb and the skeleton-free base model of a 3D character are kept in your “My library” (1.2a); the job’s progress (its steps and the services’ job numbers) stays in our jobs table. When you press “Create video”, your browser itself records the 3D stage and it goes to the video AI like a Haldun Cam recording (1.2l).

1.2oStarter → Video edit. The clip you choose from your device (or from the media bin through the AI editor’s ✦ AI assistant) (3–10 seconds; if it is longer, your browser cuts 10 seconds from the point you choose) is copied to our file store for a short time and goes from our server, together with your instruction (translated into English by Anthropic’s Claude; your own words go with it), to Google’s video AI (Gemini API). AI changes what you asked for, and the new video is saved in your “My library” (1.2a). The temporary copy in our store is deleted when the job starts (even if it fails); the file sent to Google is deleted there automatically within 48 hours. The original recording stays in your browser.

1.2pAI editor → Commercial. You write what you are advertising and what to explain, and pick videos and pictures from the media bin. When you press “📢 Make a finished commercial” or “💎 Master Tank”, your browser itself looks through the frames of each video (scenes, light, sharpness) and makes one contact sheet (JPEG) with a frame of each scene and each picture. Only the sound of the videos (16 kHz WAV, joined into one file) goes, like “AI cut” (1.2c), through a temporary copy in Supabase to the GPU and comes back as words with their times — the words are kept in your browser so the speech is not listened to again. The information you wrote, the contact sheet, the length, size and scene times of each video and picture and the words of the speech (with times) go to our server; the server matches the words against Haldun’s Mongolian dictionary and sends everything to Anthropic’s Claude. Claude sorts and splits the material and plans which parts to cut and join, the captions, title, sections, emoji, the voice-over text and voice, the AI icon and title styles, the background music direction (in English), the background behind the speaker, the relighting of very dark pictures, the overlays (logo, price sticker, lower line, effects, small window, a title behind the speaker) and the closing call to action; our server checks the plan and returns it to your browser (with the music direction encrypted). If a voice-over is needed, its text and the chosen voice go through our server to ElevenLabs and the audio comes back with the timing of each character. If background music is needed, its English direction and length go through our server to ElevenLabs and instrumental music (MP3) comes back; the music is also saved to “My library”. If a very dark picture is relit, that picture (up to 1536 px, JPEG) goes, like Relight (1.2g), through our server to Google’s image AI and a new picture comes back; the new picture is saved to “My library” and the original is not changed. If the background behind the speaker is replaced, your browser joins the spoken parts of the ad into one clip (up to 60 seconds) and sends it, like Background (1.2e), through a temporary copy in Supabase to the GPU; the clip with the new background is saved to “My library” and the original video is not changed. Your browser draws the logo, price sticker, lower line, effects and small window itself; the picture used as the logo is kept in your browser with the project. If the opening title is placed behind the speaker, your browser joins the first few seconds of the ad (usually 3–6 seconds) into one clip and sends it, like Background (1.2e), through a temporary copy in Supabase to the GPU, which turns the background green; that clip is saved to “My library”, and your browser makes the outline of the person from it without keeping it. If AI icons or an AI title are needed, the style, the icon names (up to 4; our server turns Mongolian names into English with Haldun’s dictionary) and the opening title text go through our server to Google’s image AI (Nano Banana Pro) and the finished picture comes back; your browser removes its background. Your browser edits and exports the finished commercial itself from your original videos and pictures — the originals are not sent for editing and are not changed. The finished video is saved to “My library” (up to 48 MB, Supabase) and added to the media bin as a new item (in your browser). With “💎 Master Tank”, if the commercial is 30 seconds or shorter, the video saved to “My library” also goes to the GPU service as in “Make 4K” (1.2d), and its 4K copy is saved to “My library” (the original stays as it is). Our server does not keep the contact sheet, the information, the words, the voice-over or the AI pictures (new words in the speech only as described in 1.6). The “🎬 Commercial” is made in your browser without AI — its videos, pictures and text do not go to a server; it takes no credits, and when it starts only your account and the kind (“commercial”) go to our server and are noted in its log (for sign-in and request limits) (saving the finished commercial to “My library” works as above).

1.2qAI editor → Commercial → 🍾 Sparkling. It is made from text and pictures only (no videos). Your information and contact sheet go to Anthropic’s Claude and come back as a plan, as for a Commercial (1.2p); the presenter of the voice you choose in “Voice · presenter” (or the one AI chooses) reads the voice-over. The 1–3 product pictures (up to 1024 px, JPEG) go through our server to ByteDance’s video AI (through BytePlus) and come back as a studio product shot (as in 1.2i). One product picture (up to 1536 px, JPEG) and the AI-made reference picture of that voice’s presenter (not a real person) go through our server to Google’s picture AI (or OpenAI’s, depending on our settings) and come back as one picture of the presenter holding the product; that picture is saved in your “My library” and goes from our server to Google’s video AI as the first frame of a short video. The shots are saved in your “My library” (1.2a); your browser puts the commercial together (1.2p). Our server keeps no pictures other than the shots; the English directions for the AI are kept in an encrypted token and are not sent to your browser.

1.2rAI editor → Logo motion. The logo you add does not go to our server — it stays in your browser. When you press “✨ Bring to life”, your browser draws an empty stage without the logo (the background colour you chose, with a soft glow in the logo’s colour in the middle) and sends it, together with the style, background, shape, the logo’s main colours (as colour codes) and the mood you wrote, through our server to Anthropic’s Claude, which plans the motion; with that plan and the stage picture Google’s video AI (Gemini API) makes a short video, and the stage video without the logo is saved to “My library”. Your browser then finds the hit in the video’s sound, places your real logo on the video and saves the video with the logo to “My library” (like other videos, the file does not pass through our server). The logo’s file name (or the title from Cover) is saved as the name in “My library” and does not go to the AI; our server does not keep the mood.

1.2sMNNZN — creator cells. If you place a card in a cell: the name, description, site link, cover picture (a 3:4 JPEG made by your browser) and the video, if you choose one (from your device or from My library — uploaded straight from your browser to our file store; the original in My library stays as it is), are kept in our database and a publicly readable file store; once a moderator approves the card, everyone can see it. Your email and account number are never shown on the card. The plan the card was taken with (or the trial) and your plan’s end date decide whether it is shown. We do not track people who click through to your site.

1.2tAI editor → Microphone and Remove noise. If you press “🎙 Record with microphone” and allow it in your browser, your browser uses the microphone only from the countdown until you press “■ Stop”. The recording is kept with the project in your browser (IndexedDB) on row A2 and never reaches us; if you export and save to “My library”, it becomes part of the finished video’s sound (1.2a). If you press “🎧 Remove noise”, your browser takes only the sound from the video (48 kHz mono WAV, one part per 7 minutes) and uploads it temporarily to our file store (Supabase). AI on Modal removes the background noise, makes the voice clearer and saves the clean sound (WebM) as a temporary file. The uploaded sound parts are deleted when the job ends (also if it fails); the clean sound’s temporary file is deleted as soon as your browser has downloaded and kept it. The video itself is not sent or changed, and nothing is kept on Modal. The clean sound is kept in your browser (IndexedDB); press “↺ Back to the original sound” to play the original again.

1.2uAI editor → 👤 Remove background (text behind a person). When you cut a person out of a clip copied to V2 with “⧉ Duplicate”, or out of a V2 video or picture, all the work happens on your own device: the person cutter (an open model from Google’s MediaPipe, Apache 2.0) is downloaded once from our site (haldun.ai) and makes a mask of the person in every frame. The video, the picture and the mask are not sent to anyone — the mask is kept in your browser (IndexedDB) with the project and is removed when you sign out. The tool’s original code sent usage records to Google; we removed that, so this tool does not contact any outside server. No credits or sign-in are needed. The “✨ HD · GPU” button under a V2 video is different (it uses credits): that video goes, like Background (1.2e), to the GPU service and comes back as a transparent video with no background, saved to “My library” and added to the media bin.

1.2vEditor → Reel → 💎 Reel Master. The “🎬 Reel” is made only in your browser — its text, pictures and videos do not go to a server. It takes no credits; when it starts only your account and the kind (“reel”) go to our server and are noted in its log (for sign-in and request limits). When you press “💎 Reel Master”, the template you chose, the title, text, steps, lines and closing text you wrote, the names of the places that have no picture, the voice you chose and the project’s aspect go through our server to Anthropic’s Claude and come back as Mongolian copy, a style, colour and pace, the voice-over text and voice, a music direction (in English) and picture descriptions (in English). If some places have no picture, OpenAI’s GPT Image draws up to 4 pictures in one image from those descriptions; that image is saved in your “My library” (1.2a), and your browser cuts out each picture and puts it in its place. The voice-over text and voice go through our server to ElevenLabs and the audio comes back; the English music direction and length go to ElevenLabs’ music service and instrumental music comes back, saved in “My library” as “Reel · music”. Pictures you placed yourself (screenshots) do not go to a server. Your browser adds the reel to the timeline as one clip and two sound lanes. The English directions for the music and pictures are kept in an encrypted token and are not sent to your browser; our server does not keep the copy.

1.2wHaldun AI’s Facebook Page (“Facebook Hulan”). Our server reads from Facebook (Meta) the comments written in the last 48 hours on our Facebook Page’s recent posts and reels. The comment text and the text of that post go to an AI service (Anthropic) to write a reply, which is posted publicly under your comment in our Page’s name. Your name and picture do not go to the AI. Spam, insults and political comments get no reply. While we test it, the Page’s admin sees the replies as drafts before any are posted. We do not keep or log comments, names or replies — only the comment’s number and what we did (replied, skipped) stay in our database for 30 days, so we never reply twice. You can delete your comment on Facebook at any time; to have our reply removed, send the Page a message.

1.2xEditor → ✦ AI assistant. When you give the assistant an instruction (typed, or spoken after pressing 🎤), your instruction, up to 6 earlier questions and answers, a short summary of the project (the clips, texts, sounds and layers on the timeline with their places, lengths and names, the words of your texts and the music’s beat) and a contact sheet of small frames from the videos and pictures in your media bin (one picture) go through our server to Anthropic’s Claude and come back as editing steps, which your browser carries out. When you press ✦ Start (automatic mode), no instruction or earlier questions are sent: your browser analyses the videos and pictures itself, and only the numbers (each scene’s times, light, sharpness and sound level), a contact sheet of small frames, one per scene (one picture), and the project summary go to Claude. If you speak, a recording of up to 15 seconds goes through our server to ElevenLabs’ speech recognition and comes back as text. The video, picture and sound files themselves do not go to a server. Each instruction and each automatic edit costs credits, so your account is recorded in your credit history. Videos you add as “🎨 References” are measured by your browser itself: only the numbers (shot length, motion, light, colour, tempo) and a second contact sheet with 6 small frames of each reference (one picture) go to Claude; the reference files do not go to a server, and the numbers and frames stay in your browser. When you press “💬 Captions”, only the sound of the timeline’s videos (16 kHz WAV) goes, like “AI cut” (1.2c), to the GPU and comes back as words with their times (without the AI cut) — the words are kept in your browser and the sound file is deleted when the job ends. The rows shown after “✦ Ready” go the way of their own tools: “🔆 Relight” sends dark pictures like “Relight” (1.2g), “🎧 Remove noise” sends the videos’ sound like “Remove noise” (1.2t); for “🎵 Music”, in automatic mode Claude writes an English music direction from your project, which reaches your browser only in encrypted form and, when you press it, goes through our server to ElevenLabs’ music service and comes back as music saved in “My library”. Our server does not keep or log the instruction, the reply, the analysis numbers, the contact sheet or the recording (only the time taken, whether it was automatic, the number of steps and an error code).

1.2y📞 Hulan, the phone order taker (test). When you talk to the test voice agent (the made-up “Haldun Food” menu) on ElevenLabs’ page, your speech goes to ElevenLabs and becomes text; ElevenLabs sends that text to Anthropic’s Claude to write the reply and turns the reply into speech. ElevenLabs keeps the conversation recording and transcript for 30 days — Haldun AI may listen to and read them to check the agent’s quality. When you confirm the order with “yes”, the agent sends the dishes, quantities, delivery address, phone number, extra request and the conversation number to our server. The server checks with ElevenLabs that the conversation is real, works out the total from the menu and makes a 4-digit code. The order (dishes, total, address, phone, request) is kept in Upstash for 14 days and then deleted automatically; the conversation number is not kept (only a hash for 1 day, to stop duplicate orders). Phone numbers, addresses and dishes are not written to logs. It is a test line, so no food is delivered — you do not need to give a real address or number. The agent never asks for card numbers, passwords or confirmation codes.

1.2z🎙 Talk to Hulan (haldun.ai/hulan, test). When you press 🎤 and allow it, your browser opens the microphone only while you speak (until you go quiet, up to 15 seconds) and sends the recording to our server; the microphone closes after each recording. The server sends the recording to ElevenLabs to turn it into text, then sends that text and up to 16 earlier lines of this conversation (sent by your browser) to Anthropic’s Claude to write the reply; ElevenLabs turns the reply into speech in Hulan’s voice. Recordings, transcripts and replies are not kept on our server and are not written to logs (only times and lengths); the conversation history stays only in your browser’s memory and disappears when you close the page. To protect against too many requests, daily use is counted by a hash of your IP address (1.5). If you confirm an order, the dishes, quantities, delivery address, phone number and extra request are kept in Upstash for 14 days, as with the 📞 phone test (1.2y), and then deleted automatically. It is a test, so no food is delivered — you do not need to give a real address or number. Hulan never asks for card numbers, passwords or confirmation codes.

1.2aa“🎓 Teach with my fixes” (optional). Only if you turn this switch on yourself in Upgrade (it is off by default): when you fix Mongolian‑speech subtitles with ✏ and press “🔥 Burn into video”, the audio of the part from the start up to the last line you edited by hand (16 kHz, no picture), your corrected text, the model’s original text and the line times are stored in our GPU service’s private storage (Modal in the table below), in a separate folder named by a pseudonymous hash of your account. The only purpose is improving our Mongolian speech‑to‑text model; it is not shared or sold. We record when you turned it on. Turning it off stops saving at once and deletes everything saved. Translated subtitles and dubbed videos are never included. If a recording contains other people’s voices, you are responsible for having their permission (Terms 3.12).

1.2ab“🏠 Home” page (📌 Pinterest). This page shows pictures from Haldun’s Pinterest board straight from Pinterest’s servers (pinimg.com) — when you open the page, your browser loads the pictures from Pinterest, so your IP address and browser details reach Pinterest. Our server reads the board’s list from Pinterest; we do not pass your account details to Pinterest. If you tap 📌 on a card and go to Pinterest, Pinterest’s own privacy policy applies.

1.2acMusic. The lyrics you write (up to 600 characters) and the style, voice and length you choose go from our server first to Anthropic’s Claude, which checks them for someone else’s lyrics, famous people’s names and anything harmful, and then to ElevenLabs, which makes the song. The finished song (.mp3) and the first 300 characters of the lyrics are saved in your “My library” (1.2a). If you let Hulan write the lyrics with the ✦ button, your short idea (up to 200 characters), the style and the length go to Claude, which writes the lyrics. If you have your own lyrics polished in “Lyrics fix”, your lyrics (up to 600 characters), the style and the length go to Claude and the polished lyrics come back — they are not stored, only shown on your screen. Lyrics and ideas are not written to our server logs.

1.2ad🤖 Hulan project (MNNZN, phase 1). The business description you write (up to 600 characters), the monthly budget and contact method you choose, and your phone number if you choose phone calls go from our server to Anthropic’s Claude and become a Facebook ad plan (3 ads, who to show them to, an answer list). Our server does not keep the plan; the text and phone number are not written to our logs. A picture you add with 📷 never leaves your browser — your device draws the ad pictures itself. The text and your latest plan are kept only in this browser (cleared when you sign out). If you press «🤝 Hand it to Haldun», a summary of the plan (business name, budget, audience, ad headlines), your email and phone number go to the Haldun team by Telegram — used only to contact you and run the ads. When you press «✓ Check» in «📑 Meta setup», the ad fields filled there (objective, button, campaign name, location, age, gender, interests, Page name, primary text, headline, picture text) go to Claude and are checked against Meta’s advertising rules — your phone number, website, Page ID, picture and daily budget are not sent; our server does not keep them and only the number of issues is written to our logs. The mirror is kept only in this browser (cleared when you sign out); the «📋 Step by step» copies and the downloaded picture or file are made only on your device. We never sign in to your Facebook or ask for your password.

1.2ae🔗 Hulan — Facebook connection (phase 2, test). When you press “🔗 Connect with Facebook”, Facebook’s (Meta’s) own window opens: signing in, your password and your card stay there — Haldun never sees or stores them. We work only with the Page and ad account you allow in Meta’s window. The access key (token) Meta gives us is encrypted with AES‑256 and stored in our database (Supabase) — it never goes to your browser or our logs; we also store the Page and ad account names and IDs, your Meta ID and the connection date. We use them only when you ask (for example, to save an ad to Meta). Pressing “Disconnect” revokes the permissions you gave on Meta and deletes all of this immediately. If you remove Haldun in your Facebook settings (Business Integrations, Apps and websites) or ask Meta to delete your data, Meta tells us and we delete it too; deleting your Haldun account deletes it as well. Pressing Hulan’s “🗑 Delete” (press it twice) immediately deletes the Facebook connection (revoking the permissions you gave on Meta), your business details, comments, messages and help requests from our database, and your Hulan plan and Meta setup from your browser — you can do this even after your plan has ended.

1.2af💬 ✉️ Hulan — comments and Messenger on the Page you connect (phase 2, test). Only after you set “💬 Comments” or “✉️ Messenger” to “Draft” or “On”, Meta sends new comments and messages that other people write on your connected Facebook Page to our server. While “✉️ Messenger” is on, our server also reads your Page’s latest conversations from Meta every minute (the last 5 messages of each) to find new messages — only messages that arrived after the Page’s last reply are stored. We store that text, the first word of the commenter’s Facebook name, the ID Meta gives us (needed to reply) and Hulan’s reply in our database (Supabase), and delete them automatically after 7 days. To write a reply, that text and the business details from your Hulan plan (name, products, prices, questions and answers, phone, website) go to Anthropic’s Claude. When you press “▶ Start work” (or choose a mode), these business details are stored with your Facebook connection in our database, so that “Tools” opens right away on another device too. In “Draft” mode you see the reply first and it is posted on Facebook only after you approve it; in “On” mode Hulan replies right away in the Page’s name — if a reply contains a phone number, website or email that is not in your details, it is not posted and is left for you. On Messenger Hulan replies only within 24 hours of the person’s message; if you write to that person yourself, Hulan stays out of that chat for 12 hours. If asked, Hulan says it is an AI assistant. Our logs never contain the text, names or IDs — only counts. Pressing “Off” stops new comments and messages from reaching us; pressing “Disconnect”, removing Haldun through Meta or deleting your account deletes all of this immediately. If a step fails and you press “🤝 Get help”, the step, the error message you saw, the Page and ad account names, Hulan’s modes and your email go to the Haldun team’s Telegram so that someone can contact you; the step, error code and Page name are kept for 30 days on Haldun’s admin progress board. Passwords and tokens are never sent.

1.2ag✉️ Haldun AI’s Facebook Page — Messenger. When you write to our Facebook Page in Messenger (while Hulan is switched on), our server reads new messages from Facebook (Meta) every minute. We store your text, the first word of your Facebook name, the ID Meta gives us (needed to reply) and Hulan’s reply in our database (Supabase) for 7 days, then delete them automatically. To write a reply, your text and up to the last 10 messages of that conversation (yours and our replies) go to Anthropic’s Claude; we do not store the conversation history — it is read from Facebook only at the moment of replying. Hulan’s draft reply, with your text and the first word of your name, goes to the Haldun team’s Telegram and is sent after a team member approves it; or Hulan replies right away. Hulan leaves questions it cannot answer to our team (while Hulan replies on its own, it tells you that our team will reply soon) and, if asked, says it is an AI assistant. We only reply within 24 hours of your message. Our logs never contain the text, names or IDs. You can delete your messages in Messenger at any time; to have our copy deleted, write to the Page.

1.2ah👥 Hulan — Contacts (phase 2, test). While “💬 Comments” is set to “Draft” or “On”, Hulan uses Anthropic’s Claude to tell whether a comment under a post of your connected Page shows buying interest (asking the price, wanting to order, “+”, “inbox” and similar). If it does, Hulan sends that person one Messenger message through Facebook’s (Meta’s) “private reply”, written from your business information (products, prices, how to order), and replies to the comment that the details were sent in Messenger; in “Draft” you see both first. By Meta’s rules there is only one message per comment, within 7 days of the comment, and Hulan writes to the same person at most once in 24 hours. The list of these people (“👥 Contacts”) — the first word of their Facebook name, the comment (up to 300 characters), the date, the status (sent, replied, draft, not delivered), the Messenger ID (to see who replied) and a hash of their Facebook ID (to avoid writing twice; the ID itself is not stored) — is kept in our database (Supabase) for 90 days and then deleted automatically. Only you see the list, and “⬇ Download” saves it to your own device (CSV). Logs hold numbers only — no names, comments or IDs. Pressing “✕” on a draft removes that person from the list; “Disconnect”, “🗑 Delete”, removing Haldun in Meta or deleting your account deletes the whole list at once.

1.3Sent to the server when you create. The idea you wrote, the options you chose (aspect ratio, style) and a reference picture (the first one). The server passes these to AI service providers and returns the result to you. Reference pictures are not kept; results are kept in “My library” only when you are signed in (1.2a).

1.4Technical logs. Each request creates a server log entry: date, duration, success or failure, a short error description and part of a hash of your account number. Email addresses are not written to logs. To improve quality, text that Haldun prepared from your idea (up to 4,000 characters) may be written to the log. Pictures are not logged. The hosting service keeps IP addresses and browser information in its own logs for a short time.

1.5Limit counting. To protect against too many requests and to count daily limits, we use hashes of IP addresses and account numbers (beta codes). Daily counters are deleted automatically within 2 days.

1.6New words. When you send a creation and your idea — or, when you make a commercial (1.2p), your speech — contains a Mongolian word that is new to Haldun, we keep that word with its English meaning (after an AI check) and use it to understand Mongolian text better. Only single words are kept — never your whole idea or speech, or who wrote or said it; names of people and places are not taken, and words that fail the check are recorded only as a hash.

1.7Home page sign‑up form. The email you type is kept briefly on your device and passed to the studio’s sign‑in window. It goes to the server only when you press “Get code”.

1.8We do not collect: passwords, payment card or bank account details, GPS location or phone contact lists.

2. What we use it for

We do not use your information for advertising and do not sell it to anyone.

3. Who we share it with

To run the Service, the following third‑party services process your information on our behalf:

ServicePurposeWhat is sent
Vercel (USA)site and server hostingall requests, IP address, logs
Supabase (USA)account, sign‑in and credit database; sending sign‑in codes; “My library” (storing creations); the Showcaseemail, credit history, your creations (pictures, videos, audio, ideas), published templates, public name, comments
QPay (Mongolia)taking payments for plans and extra credits (QR, bank apps) and checking themour internal invoice number, the amount and the plan name (or “Extra credits” and their number). You pay in your bank’s app — card and account details travel only between your bank and QPay. When the QR window opens, bank logos load from QPay’s server (IP address)
Anthropic — Claude (USA)🤖 Hulan — turning your business description into a Facebook ad plan and checking the ad in 📑 Meta setup (1.2ad); 💬 ✉️ writing replies to comments and messages on the Page you connect (1.2af); 👥 telling buying‑interest comments apart and writing the Messenger message (1.2ah); ✉️ writing replies in our Facebook Page’s Messenger (1.2ag); processing ideas and pictures; translating subtitles into Mongolian; turning sound effect descriptions into detailed, film-like English descriptions; turning Edit requests and Product descriptions into English; suggesting template recipes and checking templates, comments and public names; finding parts to remove in the AI editor and writing Cover ideas and Video story scripts; turning Dance shot text into English; turning Video edit instructions into English; sorting, splitting and planning the cuts, captions and voice-over of a Commercial; writing Reel Master copy; answering questions in the Haldun guide chat; writing replies to comments on our Facebook Page; writing the replies of the test phone order agent (through ElevenLabs); writing the replies in Talk to Hulan; checking the lyrics written in Music (someone else’s song, famous people’s names), writing lyrics with ✦ and polishing your own lyrics in Lyrics fixidea, reference picture, subtitle text, sound effect description, Edit request, Product description, template text, cover picture, comments, public name, speech text (word by word), guide chat questions and the page name, the song details for Cover, the Video story topic and style picture, the Dance shot actions, “Who” and “Where”, the Video edit instruction, the Commercial information, contact sheet, media lengths and scenes and the words of its speech (with times), the words you write for Reel Master, comments written on our Facebook Page and the text of that post, the text of what you said to the test agent, the text of what you said in Talk to Hulan and the earlier lines of that conversation, song lyrics, the song idea, style and length, the ad fields of 📑 Meta setup, comments and messages that reach the Page you connect, the business details of your Hulan plan, messages written to our Page in Messenger
OpenAI — GPT Image (USA)image generation, editing pictures (“Change”, “Erase” with a description), drawing covers (Cover), drawing 4 shots (Video story), the Sparkling presenter picture (depending on settings), Reel Master picturesprocessed text, reference picture, the picture to edit, the mask of the part to erase, cover ideas (in English), song title, artist, genre, mood, shot looks (in English), style picture
Google — Gemini API (USA)picture (Nano Banana Pro) and video generation, relighting pictures, editing pictures (“Change”), editing videos (“Video edit”), drawing the AI icons and title of a Commercial and relighting its dark pictures, the Sparkling presenter picture and videoprocessed text, reference picture, first frame, the picture to relight or edit, the video to edit and its instruction, the Commercial style, icon names and title text, the Sparkling product picture and the AI-made presenter reference picture, the generated picture and video
ElevenLabs (USA)voice and sound effect generation, reading Mongolian subtitles aloud, speaking the Haldun guide’s lines and answers, turning the button name you say to the guide on a phone into text, the Video story voice-over and background music, the Commercial voice-over and background music, the Reel Master voice-over and music, the test phone order agent (speech recognition, speaking the replies, keeping the conversation for 30 days), turning what you say in Talk to Hulan into text and reading the replies in Hulan’s voice, making songs in Musicthe text to be read, the sound effect description (in English), subtitle lines, the guide’s answer text, the recording of up to 3 seconds you say to the guide, voice-over lines, the music direction (in English), the generated audio, what you said to the test agent (audio, text), Talk to Hulan recordings of up to 15 seconds and the reply text, song lyrics, style, voice and length
Modal (USA)speech to text, adding subtitles to videos (including burning subtitles you corrected with ✏ again; with “🎓 Teach with my fixes” on, storing that part’s audio and your fixes for training — 1.2aa), word‑timed speech to text for the AI editor and Commercial, making videos 4K, replacing video backgrounds, restoring pictures and making them 4K, erasing the painted part of pictures, removing noise from a video’s sound (CPU), telling men’s and women’s lines apart by voice pitch for the “Male + female duo” voice-over (GPU)the audio or video file you add (to burn again with 🔥 — the original video you choose again and the corrected subtitle lines), the AI editor’s audio, videos you make 4K or replace the background of, pictures you restore, make 4K or erase from, masks, and sound you remove noise from — processed, not kept; the temporary copy in Supabase is deleted when the job ends, and the subtitled video and finished pictures are saved in “My library”, and the clean sound is deleted once your browser downloads it; voice pitch measurements are not kept and are not used to identify anyone
BytePlus — ByteDance (Southeast Asia)product video generation (AI editor → Product, the Sparkling product shot), Dance video generationproduct pictures, processed description, Dance shot text (in English), dialogue, first frame, reference pictures, a temporary link to the Haldun Cam recording or the 3D stage recording, the generated video
Tripo — VAST (Hong Kong, China)Dance → 🕺 3D characters: making a 3D character with a skeleton plus idle and walk motions from a picturethe character picture you add (made with AI or drawn by you), the generated 3D model
Uthana (USA)Dance → 🕺 3D characters: reading body motion from a video and putting it on a 3D characterthe motion video you add, the skeleton-free model of the chosen 3D character, the generated motion
Telegram (UAE)🤝 delivering Hulan “Get help” requests to the Haldun team (1.2af)the step, the error message, the Page and ad account names, Hulan’s modes, your email
Telegram (UAE)✉️ delivering drafts and “For you” notes from our Facebook Page’s Messenger to the Haldun team (1.2ag)the message text, the first word of the name, Hulan’s draft
Telegram (UAE)delivering MNNZN Studio bookings to the studiothe name, phone number or email, service, note and page address written in the form
Telegram (UAE)delivering 🤖 Hulan «Hand it to Haldun» orders to the Haldun team (1.2ad)business name, budget, audience, ad headlines, your email and phone number
Meta — Facebook (USA)reading comments on our Facebook Page and posting replies under them; ✉️ reading Messenger messages sent to our Page and sending replies (1.2ag); 🔗 Hulan — working with the Facebook Page and ad account you connect, when you ask (1.2ae); 💬 ✉️ sending us new comments and messages on your Page and posting Hulan’s replies (1.2af)the comment you already wrote on Facebook or the message you sent to our Page in Messenger (we read them), the text of our reply; the IDs of the connected Page and ad account, the data of your request; comments and messages that reach your Page, the writer’s name and ID, Hulan’s reply
Google Fonts (USA)loading fontsIP address, browser information
Pinterest (USA)the “🏠 Home” page — showing pictures from Haldun’s Pinterest boardIP address, browser information (when you open the page)
Upstash (USA), when enableddaily limit counters; test orders from the phone line and Talk to Hulan (14 days)hash of account (beta code) and IP address, counts; the dishes, total, delivery address, phone number and extra request of a phone or Talk to Hulan order
Cloudflare (USA), when enabledthe “are you human” check before a sign‑in code is sent (Turnstile), so bots cannot create many accountsIP address, technical browser and device information, the check result

Sign‑in codes are delivered by Supabase or an email service connected to it. These providers process information under their own terms. Under their API terms or our account settings, such data is not used to train their models, and it may be kept for a limited time to monitor misuse. If a new provider is added, we will update this list.

These services are located outside Mongolia, mainly in the USA, so your information is processed there. By using the Service you agree to this.

Visible to everyone. Templates published in the Showcase (the work, cover, title, description), public names, comments and heart counts can be seen by anyone on the internet. Only people who got a template see its recipe.

Bookings sent through the MNNZN Studio site go to MNNZN Studio (a sound recording studio on the Haldun AI platform), which uses them only to reply to your booking.

We may disclose information when the law requires it (an official request from a court or law enforcement).

4. How long we keep it

5. Cookies

When you sign in, we set two necessary cookies to keep you signed in: hd_at (1 hour) and hd_rt (30 days). They are HttpOnly — site scripts cannot read them, and they go only to our server. They are deleted when you press “Sign out”. We use no advertising cookies, trackers or analytics. The browser’s localStorage is used only to keep your settings and creations. If we add analytics, we will update this policy first.

6. Security

The site runs over HTTPS. AI and database keys exist only on the server. Sign‑in tokens are in HttpOnly cookies, only the server changes credit balances, logs hold hashes rather than emails, and the number of requests and sign‑in attempts is limited. Please note, though, that no method of transmission over the internet is completely secure.

7. Your rights

You have the right to know about your information, get a copy, have it corrected or deleted, and withdraw your consent. You can delete the information in your browser yourself (browser settings → delete site data). To have your account deleted, or for any other request, write from your registered email to hello@haldun.ai. If you are not satisfied with our answer, you have the right to complain to the supervisory authority under the law.

8. Children

The Service is not intended for people under 18. If you believe a child’s information has reached us, let us know — we will delete it.

9. Changes

If we update this policy, we will write the date at the top and announce important changes on the site and in the studio.

10. Contact

Haldun Ai LLC

Registration number: 7334809

Address: Building Enkhtaivan, Peace Avenue, Sukhbaatar district, Ulaanbaatar, Mongolia

Email: hello@haldun.ai