Privacy Policy — CaptionUI
Last updated: November 15, 2026
Effective: November 15, 2026
Summary
CaptionUI transcribes, translates, styles and exports captions on your iPhone. By default, your video, its audio, the transcript, the captions and the translations stay on the device, and no server of ours receives your content. There are no user accounts.
There is one exception, optional, which only happens after a yes from you: cloud transcription, a Pro feature (free accounts get 20 minutes to try). If you accept, the app sends the audio — only that of the part you chose to transcribe, never the video — through CaptionUI's server to Microsoft Azure, in the United States, which returns the text, the timing of the words and who is speaking. Neither Microsoft nor we keep the audio or the text. You can decline, or change your mind in Settings, and transcription carries on entirely on the device (section 5).
There is a second exception, also optional, with a yes of its own: AI direction and translation, a Pro feature for the directed caption styles. If you accept, the app sends the text of the words of the chosen part — never the audio, the video or the word timings — and, when you ask for a translation, the target language, through CaptionUI's server to a language model on Azure OpenAI, from Microsoft, in the United States. We do not keep the text; Microsoft does not use it to train models, but may keep it for up to 30 days only to detect abuse of the service. Without your yes, the rules on the device direct the captions (section 6).
What the app does send, whenever it is in use, is usage and diagnostics data to Datadog, our analytics and monitoring provider, in the United States: which screens you open, which steps of the app happen and how long they take, failures and crashes, device information and purchase information. That data includes nothing you filmed, said or wrote, is not linked to your identity, and is not used to track you or for advertising. There is no setting in the app to turn it off; uninstalling the app stops it.
This document exists to say that precisely — what leaves, what does not, where it goes and how long it stays.
1. Who we are
CaptionUI is an iPhone app developed by Kingdom Ideas Factory LTDA (CNPJ 45.049.427/0001-12), based in Recife, Pernambuco, Brazil.
Contact: contato@kifactory.app.
2. What we collect
There is no account. There is no sign-up, login, email or password anywhere in the app.
What is sent to Datadog. While the app is in use, it sends Datadog:
- Usage: which screens are opened and for how long, and product events with fixed names — onboarding finished, import started and its source (Photos or Files), project created, captions generated, translation completed, export started, completed or cancelled (with the resolution and format), caption files exported, subscription screen shown and closed and which feature opened it, purchase started, completed, cancelled, pending or failed, purchases restored, and cloud transcription started, completed, failed, replaced by the device or stopped by a limit (with the plan — Pro or free —, the audio's duration in seconds, how long it took, how many speakers and words, and a fixed reason code, section 5), and AI direction started, completed, failed, replaced by the rules on the device, stopped by a limit or with its translation saved (with how many words, screens, translations and links, how long it took, which model answered and a fixed reason code, section 6), and the directed styles' offer: the first project's preview card shown, previewed, used or dismissed, directed style applied, included directed video used and directed video exported with its translation (with which directed style and which bilingual arrangement, as fixed codes).
- Performance: how long each transcription stage took, how long an export took and the exported file's size in bytes, app launch time, app hangs (when the app stops responding), slow frames, long tasks and memory warnings.
- Diagnostics: the same events as the local diagnostic log (section 8) — numbers, flags and fixed codes, such as the video's duration and dimensions, how many words were recognized, the language codes detected, or an error code such as
speech.assets_missing— and crash reports with stack traces. - Device and app: device model, iOS version, app version, language and region settings, and the network connection type (Wi-Fi or cellular).
- Approximate location: country and city, estimated by Datadog from the connection's IP address. The IP address reaches Datadog, as it reaches any server a device connects to. The app does not use GPS or ask for location permission.
- Purchases: which product was bought (weekly, monthly, yearly or lifetime), the price and currency at the moment of purchase, and whether this installation is on Pro.
- A random identifier created by the Datadog SDK for this installation, used to count how many people use the app. It is not derived from anything of yours and is discarded when you uninstall the app.
What is never sent to Datadog. The video, its audio, the transcript, the captions, the translations, the names you give speakers, file names, project titles, style and preset names, anything you type, your photos, your Apple Account, your name, your email, payment details and the advertising identifier (IDFA). The app does not use Datadog's screen recording (Session Replay). Events come from a fixed list written into the app, and every field in them is a number, a yes/no or a fixed code — never something you wrote, said or filmed.
Always on. Sending happens whenever the app is in use, including while it finishes an export in the background. Without internet, events wait on the device and are sent later. There is no setting in the app to turn it off; uninstalling the app stops it and discards the identifier.
No tracking. This data is not linked to your identity, is not used to track you across other companies' apps and websites, is not sold and is not used for advertising. The app's privacy manifest declares NSPrivacyTracking = false.
Our server only receives your content if you choose cloud transcription or AI direction. Besides Datadog, the app does three things of its own: downloading the sync model, only when you ask (section 7d); only if you accept, sending the audio of the chosen part for cloud transcription (section 5); and, only if you accept, in a directed style, sending the text of the words for AI direction and translation (section 6). None of them carries your video.
3. Usage and diagnostics data: who receives it, where, for how long and why
- Provider: Datadog, Inc., a United States company, which processes this data on our behalf as our analytics and monitoring provider.
- Where it is kept: in Datadog's US5 region, in the United States. The app sends the data to
browser-intake-us5-datadoghq.com. This is an international transfer of data to the United States, made under article 33 of the LGPD (Brazil's data protection law). - For how long: up to 30 days in Datadog; after that, it is deleted.
- What for: keeping the app working (finding failures and crashes and fixing them), measuring its performance, and understanding how it is used and how it sells, to improve the product.
- Legal basis: legitimate interest (LGPD, art. 7, IX; for people in the European Union, GDPR, art. 6(1)(f)). The interest is knowing whether the app works and how to improve it; what is sent was limited to what that needs, with no content and no link to your identity.
4. What happens to your video
When you import a video, CaptionUI makes a copy of it inside the app's private area on your iPhone, so the project keeps working if the original is deleted or offloaded to iCloud. That copy, the transcript, the styles and the captions stay in that private area, which only CaptionUI can see.
Transcription happens on the device, unless you choose the cloud. The device's speech recognition uses the system's on-device mode exclusively: the app requires local processing unconditionally, and when the language model is not installed it tells you and stops, instead of sending the audio to servers. Your audio is only transmitted if you choose cloud transcription (section 5), and never as a silent substitute for recognition on the device. Caption translation, when used, works on the device, unless, in a directed style, you ask for the translation with AI direction turned on (section 6); directed styles are directed by the rules on the device, unless you choose AI direction. Rendering and exporting the video happen on the device.
5. Cloud transcription (optional)
Cloud transcription is a Pro feature; free accounts can use it for up to 20 minutes of audio in total, to try it, and the app says it is a Pro feature. It only happens after your explicit consent, asked once, the first time you transcribe, on a screen that says what is written here. You can decline ("On device only") and you can withdraw consent whenever you like in Settings › Cloud transcription. Without consent, nothing leaves the device and transcription is done on the device.
- What is sent: the audio of the part you chose to transcribe (the range kept by the trim), converted to 16 kHz mono FLAC — never the video, the images, the file name or the project title — and the project's language, when you chose one. The request carries technical proof that it comes from this app and from someone allowed to use the feature: an App Attest signature (Apple's system feature that attests the app is the genuine one, without giving access to any data of yours) and either Apple's signed receipt for your Pro subscription (it carries the original transaction identifier, the product and the dates; it carries no name, email or payment data) or, on a free account, a random identifier kept in the device's Keychain and the DeviceCheck token.
- Who receives it: the request goes to CaptionUI's server (
asr.kifactory.app, which runs on Cloudflare), which passes it to Microsoft Azure (the Azure Speech service, East US region, in the United States). Azure returns the text, the timing of each word and who is speaking, and the server passes that on to the app. - What Microsoft keeps: according to Microsoft's documentation for this mode (fast transcription), it does not retain or store the audio or the text; to tell voices apart it uses voice signals only during the request and discards them at the end, and it only tells the voices apart within the file, without recognizing who the person is.
- What our server keeps: no audio and no text — the audio stays in memory only while the request runs. It keeps only what is needed to enforce the limits: the account identifier (the original transaction's, for a Pro subscription; the random Keychain one, for a free account), the App Attest public key and counter of the device, how many minutes of audio were used (for the day, for each hour of use in the last 30 days, and in total), the times of recent requests, which devices used the account in the last 30 days, the free allowance balance and the service's daily total. Of the IP address it keeps only a scrambled version (a hash), with the times of refusals in the last hour, to block abuse. These records contain no name, email, audio or text and currently have no automatic deletion period. The server's technical log holds only counts, durations and error codes.
- DeviceCheck (Apple): on a free account, the server uses one of the two bits Apple keeps per device to remember that this device already used the free allowance — so reinstalling the app does not give the 20 minutes back. Apple does not tell us who you are or which device it is.
- Limits: on Pro, up to 30 minutes per video, 20 requests per hour, 120 minutes per day and 300 minutes in 30 days, on up to 5 devices per account in 30 days; on a free account, 20 minutes in total and up to 10 minutes per video. There is also a daily ceiling for the whole service. When a limit is reached, the app says which one and when it comes back.
- When the cloud is not used: if you say no; if you are offline; if the server is slow or fails; or if a limit is reached. Then transcription is done on the device, as always (without separating speakers), and the app says why. The switch only goes from the cloud to the device, never the other way round: nothing is sent without your yes. Filling in a part of the video later ("Transcribe it") is also done on the device.
- What Datadog receives: only counts and codes (section 2): the plan, the audio's duration in seconds, how long it took, how many speakers and words and a fixed error or limit code; never the audio, the text or the names you give speakers. Our server may also send Datadog aggregate numbers (minutes per day, refusals by type of limit, errors from the transcription service), with no account identifiers.
- Legal basis: your consent (LGPD, art. 7, I; for people in the European Union, GDPR, art. 6(1)(a)), which you can withdraw at any time in Settings, without affecting what was already done (LGPD, art. 8, § 5). This is an international transfer of data to the United States, made under article 33 of the LGPD, with your specific consent.
- The result stays with you: the text, the speakers and the captions come back to the device and stay in the project. After the 20 free minutes, what was already transcribed in the cloud stays in the project and exports as usual.
6. AI direction and translation (optional)
AI direction is a Pro feature for the directed caption styles — captions set as typography, with the word that matters made prominent — with no free allowance. A language model reads what was said and picks, for each caption, where the screen ends, which word is the key word and how strong it is; if you ask for a translation in a directed style, the same call translates each screen and points to the part of the translation that matches the key word. It only happens after your explicit consent, separate from the one for cloud transcription, asked once, when you choose a directed style, tap "Redo direction" or ask for a translation in a directed style, on a screen that says what goes and where. You can decline ("Rules on this device") and you can withdraw consent whenever you like in Settings › AI direction and translation. Without consent, nothing leaves the device and the rules on the device direct the captions.
- What is sent: only the text of the words of the part you chose (the range kept by the trim), numbered, with marks for pauses and sentence ends, and — only when you ask for a translation — the target language, as a standard code such as
pt-BR. Never the audio, the video, the word timings, the project title, the file name or speakers' names. The request carries the same technical proof as cloud transcription on a Pro subscription: the App Attest signature and Apple's signed receipt for your Pro subscription (section 5). - Who receives it: the request goes to CaptionUI's server (
asr.kifactory.app, which runs on Cloudflare), which passes it to Azure OpenAI, Microsoft's service that hosts language models, in the same Microsoft Azure resource section 5 names, in the United States. The model returns, for each screen, where it ends, the key words and the strength and, if asked, the translation; the server checks the answer and passes it to the app, which checks it again. - What Microsoft keeps: according to Microsoft's documentation for Azure OpenAI, the text sent and the answer are not used to train models and are not passed to OpenAI. To detect and investigate abuse of the service, Microsoft may keep the text and the answer for up to 30 days, and then deletes them.
- What our server keeps: no text and no translation — the request stays in memory only while it is served. To enforce the limits it keeps, alongside the account records in section 5, how many directions the account asked for that day and how many words it directed that month, and the whole service's daily word total; these records contain no name, email or text. The server's technical log holds only counts, which model answered and error codes.
- Limits: on Pro, up to 40 directions per day and up to 6,000 words per direction (about 30 minutes of speech). There is also a daily word ceiling for the whole service. When a limit is reached, the app says which one and, where there is one, when it comes back.
- When AI is not used: if you say no; if you are not on Pro; if you are offline; if the server is slow or fails; or if a limit is reached. Then the rules on the device direct the captions, sending nothing, and the editor says which director made the captions. A screen whose answer fails the check is also directed by the rules on the device, that screen alone. The direction is saved in the project: reopening the project or exporting again does not call the model.
- What Datadog receives: only counts and codes (section 2): how many words, screens, translations and links, how many screens stayed with the rules, how long it took, which of the models answered and a fixed error, reason or limit code; never the text, the translation or the language. Our server may also send Datadog aggregate numbers (requests and words per model, screens that stayed with the rules, errors from the AI service), with no account identifiers.
- Legal basis: your consent (LGPD, art. 7, I; for people in the European Union, GDPR, art. 6(1)(a)), which you can withdraw at any time in Settings, without affecting what was already done (LGPD, art. 8, § 5). This is an international transfer of data to the United States, made under article 33 of the LGPD, with your specific consent.
- The result stays with you: the screens, the key words and the translation come back to the device and are saved in the project.
7. What else leaves the device
Besides the usage and diagnostics data in sections 2 and 3, the optional cloud transcription in section 5 and the optional AI direction and translation in section 6, there is other network traffic and there are copies going off the device that a generic text usually leaves out. None of it carries your content to us, but you deserve to know about each one:
a) Language model downloads (Apple). The first time you transcribe or translate in a language whose model is not yet installed, the system downloads that model from Apple's servers. It is a download: none of your content goes up with it. The download is performed by the operating system, under Apple's privacy policy.
b) Purchases (Apple). If you subscribe to Pro, the purchase is handled entirely by Apple. CaptionUI does not see, ask for or store any payment data. The device keeps only a flag saying the subscription is active; Datadog receives the product, price, currency and outcome of the purchase (section 2), never payment details.
c) iOS backup. CaptionUI's projects live in the app area that iOS includes in the default backup. If you have iCloud Backup turned on, a copy of your imported videos and your transcripts may be sent to iCloud by the system, under Apple's privacy policy and your own settings. CaptionUI neither performs nor controls that upload — but saying "nothing leaves the device" without mentioning it would be inaccurate. You can turn off backup for the app in Settings › [your name] › iCloud › Manage Storage.
d) The sync model (optional). If you choose to download the model that sharpens karaoke caption timing, the app fetches that file — about 277 MB, which take up about 317 MB once unpacked — from a download server we operate (models.kifactory.app). It happens only if you tap "Download": never alongside installing the app, never automatically, never in the background. The request carries none of your data: no body, no identifier, nothing about you or your videos. Like any web request, it shows that server your IP address. The file is checked against a SHA-256 checksum before it is used, and you can remove it afterwards in Settings › Models to reclaim the space. The app works without it — every caption style works, with word-level rather than syllable-level precision.
e) What you share yourself. Exported videos, caption files (SRT/VTT/ASS), style presets and the diagnostic log only leave the device when you use the share sheet or save to the Photo Library. You pick the destination, and the destination receives the file. If you save to the Photo Library and have iCloud Photos enabled, the video enters that sync.
8. Diagnostic log
CaptionUI keeps a local technical log so you can report problems. It is structurally incapable of containing what you said or wrote: the format has no free-text field. Each event carries only numbers, flags and fixed codes — for example, how many words were recognized, how long an export took, or an error code such as speech.assets_missing. There is no transcript, no file name, no path.
The same events, with the same fields and nothing more, are also sent to Datadog automatically (sections 2 and 3). The local copy stays on your device and only leaves it if you tap "Share" on the Diagnostics screen and choose where to send it. You can delete the local copy at any time with "Clear log", on the same screen; that does not delete what was already sent to Datadog, which is deleted at the end of the up-to-30-day period in section 3.
9. Permissions
CaptionUI asks for two permissions, both only when they are needed:
- Speech recognition — to transcribe your video's audio on this device.
- Add to Photo Library — only to save the captioned video when you ask for it. It is an add permission, not a read one: the app has no access to your photos.
Cloud transcription (section 5) and AI direction (section 6) ask for no permission: App Attest and DeviceCheck are system features that show no screen and give no access to your data.
We do not ask for camera, microphone, location, contacts, or tracking permission. The approximate location in section 2 is estimated from the IP address, not from the device.
10. Children
CaptionUI is not directed at children. The usage and diagnostics data in section 2 is sent by every installation of the app, without knowing the user's age and without identifying the person: we do not ask for or keep a name, email, age or any other identifying information.
11. Your rights (LGPD / GDPR)
You have the rights set out in the LGPD (art. 18) and, if you are in the European Union, in the GDPR — among them, to confirm whether we process data about you, to access it, to ask for it to be deleted, and to object to processing based on legitimate interest.
In practice, because the data sent to Datadog is not tied to a name, email or account, we generally cannot single out one person's records. You can still write to contato@kifactory.app and we will answer your request. Uninstalling the app stops the collection, and what was already sent is deleted within 30 days.
Your content — videos, transcripts, captions — stays on your device and under your control: deleting a project in the app, or uninstalling the app, removes it.
Cloud transcription. The audio and the text stay neither with us nor with Microsoft, and what our server keeps (section 5) is not tied to a name or email: the account identifier is an Apple number, on Pro, or a random device identifier, on a free account. So here too we generally cannot single out one person's records; write to contato@kifactory.app and we will answer your request. Withdrawing consent in Settings stops any further sending.
AI direction and translation. The text does not stay with us. The request that reaches Microsoft, which it may keep for up to 30 days to detect abuse (section 6), carries no identifier of yours; our server's counters use the same Pro account identifier as section 5. So here too we generally cannot single out one person's records; write to contato@kifactory.app and we will answer your request. Withdrawing consent in Settings stops any further sending.
For the parts described in section 7 that are operated by Apple (model downloads, purchases, iCloud backup), Apple's privacy policy and controls apply.
12. Changes to this policy
If this policy changes, the "last updated" date above changes with it, and the new version becomes available at the same address and inside the app.