Privacy Policy

1Who we are and how to reach us

METAG.AI operates the website at metag.ai, the browser-based Studio editor, the macOS application, and an MCP endpoint that lets external AI agents act on your account. For any privacy question, request or complaint, write to hi@metag.ai. A person answers.

2What stays on your device

The following runs entirely on your own computer. We never receive the media itself.

Speech transcription

In the browser, transcription runs inside a Web Worker using Whisper via transformers.js (model onnx-community/whisper-base_timestamped), producing word-level timestamps that drive our karaoke-style captions. On macOS, transcription runs through Apple's SpeechAnalyzer framework. In both cases the audio is processed locally and is not sent to us. The browser does download the Whisper model weights directly from the Hugging Face CDN, which means Hugging Face sees your IP address and which model you requested. It does not see your audio.

Media analysis

Loudness analysis, keyframe extraction and thumbnail generation run locally.

Editing

Timeline editing, effects, captions, region-based watermark removal and export are performed locally. In the macOS application this covers the entire editing and export pipeline.

3What leaves your device

This is the complete list we are aware of. Three of these items are your source material itself rather than a description of it, and we call them out for that reason.

Text you write

Prompts, topics, storyboards, narration text, editing instructions and preset choices are sent to our gateway and from there to a large language model provider, in order to plan shots and write narration. Our default text model provider is Alibaba Cloud Model Studio (DashScope). We store the prompt and the storyboard in our database, subject to the retention rule below.

Transcripts and loudness data

When you use the editor's AI assistant or the highlight-finding feature, the transcript produced on your device and the loudness curve derived from your audio are sent to our gateway and forwarded to a language model provider. The transcript is the words spoken in your footage. We do not store it in our database, but it does leave your device.

Keyframe thumbnails — your footage, as images

When you ask the editor's AI assistant to work from what it can see, your device extracts still frames from your own footage and sends them, as images, to a vision language model provider (by default Alibaba Cloud Model Studio's qwen-vl-max). Up to forty-eight frames may travel in a single request. These are pictures of your material, not a text summary of it. We do not store them.

A voice sample — your recorded voice

Voice cloning requires uploading an audio sample of a voice. It gets its own section below.

An image you ask us to edit

The image-editing feature requires uploading that image. It is stored in our object storage under a key derived from a hash of your account identifier, made reachable by the image model provider through a time-limited signed link, and edited by Alibaba Cloud Model Studio's qwen-image-edit model.

Narration synthesis

By default, narration text for generated films is synthesised by Alibaba Cloud Model Studio's CosyVoice, so the narration text leaves our servers. A local synthesis engine runs as a fallback when the cloud call fails, and is what serves the free in-editor speech synthesis when you have not selected a cloned voice.

Websites you point us at

The website-to-promo preset makes our server fetch the URL you supply and run a headless browser against it to take screenshots. The target site sees a request from our server, identified by the user agent METAG-SitePromo, not from you.

Cloud generation

If you choose a cloud engine tier, the shot prompts and any first-frame image are sent to that provider to render the shot. Depending on the tier this is Alibaba Cloud Model Studio, Google (Gemini and Veo), xAI (Grok Imagine), or Volcengine (Ark and Seedance). The in-house engine tier renders on our own hardware.

Payment details

Card and billing data go directly to Stripe. We never see or store your card number.

If you supply your own API key

The Studio can call certain third-party services directly using a key you provide and store encrypted in your own browser. One of these, KieAI image generation, uploads the image you give it to KieAI's servers; their own documentation states that generated media files are deleted after fourteen days and logs after two months. When you use a bring-your-own-key feature, that provider's terms govern what happens to your data, not ours.

4Third-party resources our apps load

Separately from anything you create, our web pages and applications fetch code, fonts and model weights from third-party hosts. Those hosts see your IP address, your browser or device details, and which resource you asked for. They do not see your media. We list them because "your footage stays on your device" is a statement about your footage, not a claim that your browser talks to nobody.

Our website and the Studio editor load web fonts from Google Fonts on every page load. The Studio additionally fetches, when the corresponding feature is used: Whisper model weights from the Hugging Face CDN, MediaPipe segmentation models from Google's storage CDN, MediaPipe and FFmpeg builds from unpkg and jsDelivr, a media-decoding library from esm.sh, and video stabilisation binaries from mediashares.openreel.video. The macOS application downloads Apple's speech model through Apple's own asset service, downloads a visual-search model from Hugging Face, and checks for updates against our update feed.

5Data stored on your own device

The Studio keeps your projects, your imported media, autosave snapshots, templates, fonts, and the transcripts and loudness curves it derived from your media in your browser's IndexedDB. Any third-party API keys you enter are stored encrypted, under a password you set, in the same place. Your sign-in token and your language preference are kept in browser local storage. All of this is on your machine and we cannot read it. Clearing your browser's site data removes it.

6Voice cloning

A voiceprint can identify a specific human being. Several jurisdictions treat it as biometric or personality-right data requiring explicit consent, and our code already treats it that way. This section exists so that the treatment is visible to you and not only to our engineers.

Consent is a precondition

Our API rejects a cloning request that does not carry an explicit confirmation that the sample is your own voice or that you hold the rights to it. No path through the product clones a voice without that confirmation.

What we record

For each cloned voice we store a name you choose, the identifier the provider assigned to the voice, the model it was registered against, the timestamp of your consent, and the IP address from which consent was given. The timestamp and IP exist so that a later dispute about whose voice this is can be settled on evidence rather than on assertion.

Where the sample goes

The uploaded sample is written to our object storage, made reachable by Alibaba Cloud Model Studio through a signed link valid for seven days, and enrolled with their CosyVoice service. The provider then holds a derived voice model on their systems until it is deleted.

Deleting a voice

Deleting a cloned voice removes our record and instructs the provider to delete the voice on their side. Our integration treats a failed provider deletion as a failure rather than reporting success, so that we never tell you a voice was removed when it was not.

A limitation we are disclosing rather than hiding

As of this draft, deleting a single cloned voice does not delete the original audio sample from our object storage, and our account-deletion routine does not yet include voice samples among the objects it purges. This is a defect, we are fixing it, and until the fix ships you should assume a voice sample persists in our storage. If you want one erased now, email hi@metag.ai and we will remove it manually and confirm that we have. We would rather write this down than let you infer a guarantee we do not currently keep.

7Account and identity data

You sign in through Google, GitHub or Apple. We receive and store an account identifier of the form provider:id, plus your email address when the provider tells us it is verified. We do not store your avatar or your display name. Alongside your account we hold your credit balance, your subscription expiry, your Stripe customer identifier once you have paid, and the time of your last sign-in.

Sign-in log

Every sign-in writes a row recording the account identifier, the provider used, the email address supplied, the source IP address, and whether that sign-in created a new account. The log exists to resolve account-takeover claims and to identify duplicate accounts when one person has signed in through two different providers.

Free-credit deduplication

New accounts receive a one-time credit grant. To stop one person collecting it repeatedly through different sign-in providers, we record that a grant was made against a salted SHA-256 hash of the verified email address. We store the hash, not the address. It cannot be read back as an email address, and it is deliberately retained when you delete your account, because otherwise deleting and re-registering would be an unlimited source of free credits. We also count new accounts per source IP address in a short-lived cache in order to detect bulk signups.

8Advertising attribution

We run no third-party analytics or advertising script on our website. There is no Google Analytics, no Mixpanel, no tag manager and no advertising pixel on metag.ai. An earlier version of this policy claimed otherwise; that claim was inherited from a template and was never true of this product.

What we do instead happens on our servers. If you arrive from a Google Ads click, the click identifier and any campaign parameters travel with your sign-in request and, only if that sign-in creates a new account, are stored once on your account record. They are never overwritten afterwards. We use them to report back to Google Ads which clicks resulted in a signup or a purchase, so that we can tell which advertising is worth paying for. The report we upload carries the click identifier, a timestamp and a conversion value. It does not carry your email address and it does not carry anything you created.

9Product analytics and crash reporting

In the browser

The Studio editor contains code that can send product events — such as "project created" or "export finished" — to PostHog, together with page views. It activates only when a PostHog key is configured at build time. Event names and counts are sent; nothing you create is. We will state whether it is switched on, here, before these terms take effect.

In the macOS application

Builds compiled with telemetry enabled include crash reporting through Sentry and product analytics through PostHog. Both are on by default and both can be turned off in the application's settings. Crash reports are configured not to attach personally identifying request data. The analytics layer sends only an explicit allow-list of events (application opened, project created, export started, finished or failed, agent session started, agent tool called, MCP session activated) and strips any property not on that list; screen views and automatic lifecycle capture are disabled. Neither carries your media, your prompts or your transcripts.

macOS update checks and download counter

The macOS application periodically checks our update feed, which reveals your IP address and installed version. Downloading the application from our website records the version, the browser user agent and the referring page; that record is not linked to any account.

10How long we keep things

Generated files

Rendered shots and audio are written to our object storage and kept there, so that a film you paid for stays retrievable. They are deleted when you delete the film or close your account. A short-lived copy also sits on a RAM-backed filesystem on the machine that produced it and is cleared within a few hours; that copy is a cache, not the record. This changed on 1 August 2026. Until then the rendered files existed only in RAM and were deleted about an hour after they were produced — which meant a job could report success while the files behind it were already gone, and a user could be charged for a film that no longer existed. We would rather hold your work than bill you for something that disappears. The creative inputs are still erased on the schedule described next.

Job metadata

Each generation job carries an expiry set twenty-four hours after it was created. A sweeper then blanks the prompt, the storyboard, the first-frame reference and any error message on that job. It does not delete the row: the engine used, the credits charged and our own platform cost are kept so that we can reconcile our books. After twenty-four hours, in other words, we retain the accounting and not the creative content.

Uploaded images and voice samples

These are held in object storage until you delete your account, subject to the limitation described in the voice-cloning section.

Product events

When you export a film we record that an export happened, how many clips were on the timeline, how long it ran, and whether any of the shots were generated by us. We record no part of the content itself — no prompt, no narration, no frames, no file. We added this on 2 August 2026 because we had no way to tell whether anyone ever exported a finished film, and a product nobody exports from has not delivered anything. You can ask us to delete these rows at any time by emailing hi@metag.ai.

Account, billing and audit records

Your account row, credit ledger, payment events and sign-in log persist for as long as your account exists. Payment events and the sign-in log are retained after account deletion; the next section says why and what you can do about it.

11Deleting your account

You can delete your account from within the product. The request must carry an explicit confirmation, because the operation is irreversible. When you make it, in this order, we do the following.

First, we instruct the voice provider to delete every cloned voice you registered. This happens before we delete our own records, because once our records are gone we no longer know which voice to delete on their side. A failed provider deletion is logged for manual follow-up rather than being silently swallowed.

Second, we delete your user row, which cascades to your generation jobs, your cloned-voice records and your auto-director runs, and we separately delete your credit ledger.

Third, we delete your files from object storage one at a time, verifying each deletion with a follow-up read. We do this because bulk deletion on our storage provider has been observed to report success without actually deleting, and we are not willing to tell you a file is gone on the strength of a response we have not checked.

Fourth, we clear the transient rate-limit and upload-ownership entries held for you in our cache.

What survives deletion

Three things do not go away, and you are entitled to know which. The salted hash of your email address in the free-credit register stays by design, so that deletion cannot be used to farm free credits; it is a hash, not an address. Payment event records stay, because records of money received have to be kept. The sign-in log stays, and it contains your email address and IP addresses; we have not yet documented a retention decision that justifies keeping it, and we are reviewing whether to purge it on deletion. If you want your sign-in log purged now, email hi@metag.ai and we will do it.

12We do not train on your content

We do not use your prompts, your footage, your transcripts, your voice samples or your generated outputs to train, fine-tune or evaluate any model of ours. We operate no training pipeline that consumes customer data. What third-party model providers do with material submitted to them is governed by their own terms; we select providers with that in mind, but we cannot make the promise on their behalf.

13Who else sees your data

We share data only with the providers required to run the features you actually use. As of this date those are Alibaba Cloud Model Studio (text models, vision models, image editing, speech synthesis, voice cloning), Google (Gemini and Veo generation, and Google Ads conversion reporting), xAI (Grok Imagine generation), Volcengine (Ark and Seedance generation), Cloudflare R2 and Alibaba Cloud OSS (object storage), Stripe (payments), Hugging Face and the other content delivery networks named above (delivery of model weights, fonts and code to your device), PostHog and Sentry (product analytics and crash reporting, where enabled), and whichever sign-in provider you chose. If you use a bring-your-own-key feature, that provider — for example KieAI, ElevenLabs, OpenAI or Anthropic — receives what you send it, under your own account with them. We do not sell your data and we do not share it with advertisers or data brokers.

These providers operate in several countries, including mainland China and the United States, so using the service means your data may be processed there. We have not yet completed a formal cross-border transfer assessment. We expect legal review to require one.

14Your rights

Depending on where you live you may have the right to access, correct, export or delete your personal data, to object to or restrict processing, and to withdraw consent. Deletion is available directly in the product. For anything else, write to hi@metag.ai and we will act on it. We will not charge you for exercising a right and we will not degrade your service because you exercised one.

15Cookies and browser storage

We use browser storage to hold your sign-in token and your language preference. We use no advertising or cross-site tracking cookies.

16Children

The service is not intended for anyone under eighteen. We do not knowingly collect data from children. If you believe a child holds an account, tell us and we will delete it.

17Security

Generated files live only in memory-backed storage and expire. Access to a job's files requires a short-lived ticket scoped to that single job, rather than a long-lived token sitting in a URL. Object storage keys are namespaced by a hash of your account identifier rather than by anything readable. We do not claim our systems cannot be breached; nobody honest does.

18Changes to this policy

We will post material changes here with a new date and keep the history below, so that you can see what changed and when.

July 29, 2026

First version written for METAG.AI. It replaces an inherited template that described another company's product, named analytics providers we have never used, and carried an effective date predating this company. This version adds the on-device processing description, the explicit list of what leaves your device, the voice-cloning section, retention windows, the account-deletion procedure and the advertising attribution disclosure.