Privacy
Privacy at Takehound
Last updated November 17, 2026.
Most of this page is about the editor at /editor, where Hound works. The older tool at /app and the studio come after it, under their own headings.
The editor at /editor
This editor runs in your browser, and keeps a copy of your projects and your video files in your browser's own storage, the video files in one footage library that every project uses. If you have an account, Takehound's server keeps the copy of record: every project in your browser goes to your account, not only the one you open, when the editor or its list of projects loads and when you sign in, so it is there on any computer you sign in on and is not lost if this browser's data is cleared. A project made while you were not signed in goes to your account when you sign in. The video files are not uploaded unless you have an account and Takehound's footage storage is on (see Your footage library, below); then a file that a project kept apart from the library is moved into the library first, so it is uploaded like the rest. When the editor cleans the sound of your recording, and when it enhances your voice, it does that in your browser too and keeps the cleaned copy in your browser's own storage beside your video files; nothing of the sound is sent for it, and if you have an account, only how strongly the sound was cleaned and whether your voice was enhanced go with the cut. When a card takes its colour from your video, your browser reads a few small frames for it and sends none of them: only the few colours it finds, as colour values, are kept with your project, go with the cut if you have an account, and are named to the editor's model so it can choose one. To clean it, your browser downloads a noise model (DeepFilterNet3, about 20 MB) from Takehound's own server the first time, and asks for it again when its copy is a day old; nothing of yours goes with those requests. Six things are sent. Five go when you press their button; one of them, Describe my footage, also goes when you press Do the whole first pass while files are waiting to be described, and one, Transcribe, also goes on its own for some recordings you add with an account (see below); the sixth, the review of your cut, goes after a first pass and when you ask the editor to check your cut (see The review below). Transcribe sends the sound of your recording to Takehound's server, which turns it into words and deletes the sound once it has been read. If the recording is in your footage library with your account (see Your footage library), the server reads its sound from the copy kept there instead, and nothing is sent from your browser for it; the words, and how loud the recording is moment by moment, are then kept with the clip. Since 8 November 2026, if you are signed in, a recording you add with one of the add buttons in Takehound's panel in the editor, or by a drop on the panel, to a project with nothing yet on its timeline's main video track is transcribed with no press once it has finished uploading to your footage library, when it has no words yet, the editor has read its length from the file, and it is not longer than a free video takes, whatever your plan: the server reads its sound from the copy in your footage library, as above, and your browser does not send its sound for it. While it is still uploading, the panel says it will be transcribed as soon as the upload finishes, and its Cancel stops that. A recording whose length the editor cannot read from its file, or one longer than a free video takes, is not transcribed without a press, whatever your plan. If the server cannot read the library's copy (footage storage is off, or the upload fails), nothing is sent for it until you press Transcribe. Describe my footage sends a few small stills of each clip to Takehound's server, which passes them to Anthropic's API to be described and does not keep them, the same way the tool does. For a clip in your footage library with your account, the server takes those stills from the copy kept there instead of your browser sending them; they go to Anthropic's API the same way and are not kept, and what each clip shows is kept with it. Since 13 November 2026, pressing Do the whole first pass while clips or pictures beside your recording have no description yet first describes them exactly as Describe my footage does: a few small stills of each go to Takehound's server and on to Anthropic's API to be described, or, for a clip in your footage library with your account, the server takes them from the copy kept there, as above. The first-pass card says so while files are waiting. The pass is made once they are described; Cancel while they are being described stops the press, and no pass is made. Asking the editor for something sends the words of your recording, a description of your timeline (file names, times, text on screen), your notes and what you told it about your footage through Takehound's server to a model. While it works the editor also looks at the picture: it sends a few small stills of your video as it stands, each time it checks a moment, the same way. What you ask is done by Hound's core whenever Takehound's server can run it, and sends what that core's jobs send (see Hound's core). When it cannot, it is done the usual way. For a lesson, the model is Anthropic's, through its API. For a story's asks, it is Google's Gemini model, and the editor also watches and listens: a small copy of the cut, picture and sound at 480p, goes with the rest through Takehound's server to Google's Gemini model through OpenRouter, which relays it, with zero data retention, and it is not used to train a model. Takehound's server holds that copy in memory for that job only, at most two hours, and never on disk. If Gemini cannot be used for a reason on Takehound's side, the job goes on through Anthropic's API instead, and the copy of the cut goes no further. On a story, Do the whole first pass is made by Hound's core whenever Takehound's server can run it, and sends what that core's jobs send (see Hound's core). Otherwise Do the whole first pass sends the words of your recording with their times, which of them are already cut, the file name and the description of each of your clips, and what you told the editor about your footage to Takehound's server. The server passes the words, the descriptions and what you told the editor on to a decision model (see The decision model below), which picks the cuts, the lines and the clips from them; your file names and the times go no further than Takehound's server, and no picture and no sound goes to the decision model. The editor makes the pass that way when it can. When it cannot (the video is a lesson, whose slides have to be looked at; the editor has written down how you cut; the video already has things on screen; or the decision model is off, busy or cannot be reached), the press is a job like any other you ask the editor for and is sent to Anthropic's API as a lesson's asks are, stills of your video included; if the decision model was asked first and failed part way, the words had already gone to it. Takehound's server does not keep any of it. Picking a font fetches that font from Google Fonts, and the Captions tab downloads a speech model from Hugging Face and transcribes in your browser. Grading the colour of your recording decodes a few small frames of it in your browser, and what they measure as (how bright the picture is, how deep its blacks go, its colour cast, the colour of the parts of it that ought to be white or grey, and how strong its colours are) is kept with the cut; no picture is sent. Pushing in at the cuts of a lesson decodes a small frame of its recording for each second in your browser, or, for a recording in your footage library with your account, on Takehound's server from the copy kept there, and where the recording shows a slide of its own (the seconds each starts and ends, so no push-in crops it) is kept with the cut, and with the clip in your library; no picture is sent anywhere else. Commands the editor learns are kept in your browser. If you have an account, they and each cut (the timeline, its reasons and notes, the words, your file names) are kept with it too; the video files are kept with it only while footage storage is on (see Your footage library). Your settings in the editor (whether a review follows a first pass, which view of the editor shows, the switch between Hound's two cores, your keyboard shortcuts, your easing presets and the sounds you saved) are kept in your browser and, if you have an account, with it, each with when you chose it, so another computer you sign in on uses them too.
The editor learns how you cut, from what you do with its work and from what you tell it. It keeps a short list of that in your browser. Part of the list is worked out from what you do, and is kept with what it was worked out from: which of the editor's steps you kept and which you took back, the numbers you kept instead, the id of the project each was seen in, and how many of its steps you took back in each recent job. The rest is sentences about how you like your videos cut: ones the editor wrote down, each with its reason and when it was written, and any you write there yourself. A sentence can carry what you said to the editor. Along with the list, the editor keeps the look you chose last for what goes on screen: Bold or Clean, your own colour if you gave one, and when you chose it. Your next video starts in it, and asking the editor for something sends the look of the open video along, with your colour if you gave one. You change it in the Takehound panel under Look. It keeps the colour grade you chose last for your recording the same way: Natural, Bold, Clean or none, how strong, and when you chose it, and your next video is graded by it. It also keeps how strongly you chose to clean the sound of your recording and whether you chose to have your voice enhanced, or that you turned the cleaning off, and when you chose it, and the sound of your next video goes by it. When the editor suggests next steps after a first pass and you turn one down, it keeps which kind of step that was, the id of the project and when, and once you have turned a kind down on two videos it does not suggest it again. When Hound, the dinosaur in the editor, cheers one of your firsts (your first finished pass, the first video or timeline file you make, every fix of a review made, the hum taken off your voice), it keeps which first that was and when, with the list and with your account if you have one, so that it cheers each one only once; clearing the list clears that too.
Asking the editor for something sends the list itself along, through Takehound's server to Anthropic's API or to Gemini like the rest, and not what it was worked out from; Takehound's server does not keep it. You can read the list and change it in the Takehound panel under What the editor has learned about how you cut: put a line, the grade or the sound you chose, or a kind of next step it no longer suggests, back to the usual, reword a sentence the editor goes by or take it out, say yes or no to one it noticed once, or clear all of it. A sentence that is taken out, by you or by the editor, is still kept for a time, and the latest of them are still sent, marked as taken out, so that the editor does not write them down again; they show at the end of the list, and clearing the list removes them too. The feed of a cut also shows these sentences, each time the editor writes one down, changes it or takes it out, and when a job starts from them, and those lines of the feed stay with that cut when the list is cleared. If you have an account, the list, what it was worked out from and the look, the grade and the sound you chose are kept with it too.
Your YouTube post
Draft my YouTube post, and asking the editor for a title or a description, is a job like any other you ask for and sends what such a job sends (see Hound's core), with the links and the line you keep for your descriptions. What it drafts, the title options and the description, is kept beside your cut: in your browser, and with your account if you have one. The links and the line you type for the end of your descriptions are kept with the list of how you cut, in your browser and with your account if you have one, so your next video has them. Forget all of it leaves them, because you wrote them. Empty the box to take them out, and Delete this account removes them. Nothing is sent to YouTube: you copy the post and paste it there yourself.
Your YouTube channel
Under Your YouTube channel in the Takehound panel you can bring the files YouTube Studio exports for you: the table of your videos (titles, when each was published, length, views, impressions, click-through rate, average view duration, percentage viewed, subscribers) and a video's audience retention. They are read in your browser and kept there and, if you have an account, with it: Takehound's server keeps the numbers you brought (in Cloudflare R2 while footage storage is on, otherwise in its own database), so they are there on any computer you sign in on. Clear takes them off this computer and out of your account, and Delete this account removes them. Takehound does not connect to your YouTube account to get them. When the editor reads them for a job you ask for, the numbers it reads go along with that job, the same way as the rest of what a job sends (see Hound's core). What the editor learns from them is written in the list of how you cut, marked as measured on your channel, and kept there like any other line (see above): you can read it, reword it or take it out.
Your footage library
In the editor, every file you add to a project goes into one footage library in your browser, and every project uses it, so a clip is added once and placed in any project. What is known of each clip (what it shows, and its words if it was transcribed) is kept beside it in your browser and made once. Asking the editor for something can send the names and descriptions of the clips in your library that are not in the open project yet, the same way as those of the clips in it. When your library is kept with your account, the editor's search of it for a clip that fits a line runs on Takehound's server, over what it keeps of your clips: their names, what they show and their words. Only the words searched for are sent for it. The library says whether your clips are kept on this computer only.
If you are signed in and Takehound's footage storage is on (the library says so), your video files are uploaded and kept in Cloudflare R2, a storage service run by Cloudflare, a separate company, so your library is there on any computer you sign in on. Your browser sends each file straight to R2. With your account, Takehound keeps each clip's name, size, length and picture size, a small picture of it, what it shows, its words if it was transcribed, how loud it is moment by moment (kept in R2 beside the file), where a recording shows a slide of its own, a copy of its name, what it shows and its words arranged for searching your library, and where its file is in R2. When you press Transcribe or Describe my footage, when you press Do the whole first pass while a clip has no description yet, when a recording you add is transcribed on its own (see The editor at /editor), or when you push in at the cuts of a lesson, Takehound's server reads the clip's file from R2 to do it, on Takehound's own server. It reads the file as it goes and keeps nothing of it on the server; only what it found is kept, with the clip. On another computer a clip plays from R2 through an address that works for an hour, and putting it in a project there copies it into that browser's storage. An account's library can take up to 50 GB.
Your video files and what is known of them are kept until you delete them. Remove, on a clip in the library, deletes it from your library: its file in R2 and its loudness beside it, what Takehound keeps of it, and its copy in that browser; the projects that used it then say it is missing. Delete this account, on the account page, deletes every file of your library in R2 with the rest of the account. With footage storage off, or when you are not signed in, nothing of the library is uploaded.
Accounts
Trying Takehound starts with signing in with your Google account (since 18 October 2026). Takehound stores a random id, the id Google gives your Google account, its email address when Google says it is verified, the name on it when the account is made that way, your projects, the list of how you cut that the editor keeps, the look, the colour grade and the sound cleaning you chose, and your footage logs (the description of each shot and your file names). With the editor it also stores every project in your browser, Hound's conversation on each project, your settings in the editor, and your YouTube channel's exports if you brought them (see The editor at /editor, Hound's core and Your YouTube channel). While footage storage is on, it keeps your footage library from the editor too, your video files included (see Your footage library). A template you publish is kept with it too (see Templates). Delete this account, on the account page, removes all of it, projects and templates included. It deletes the files of your footage library too.
When a Google sign-in makes a new account (since 8 November 2026), Takehound also keeps where the visit that led to it came from, as your browser tab noted it when you arrived (see How we count visits): the utm_source, utm_medium, utm_campaign and utm_content of the link you landed on, if it had them, the host name of the site that sent you, if one did (only the host, such as www.reddit.com, never the rest of that page's address), the path of the first page you landed on here (since 9 November 2026: only the path, such as /start/experts, never what came after it in the address), and when you arrived (or, if the tab did not say or it was more than 30 days before, when the account was made). It is carried from that first page to the sign-in in the tab's own storage, with no cookie. It is kept once, when the account is made, and never changed by a later sign-in. Kyle uses it to learn which places bring people who go on to use Takehound. Only Takehound's owner sees it, in the owner's list of accounts and in the copy of an account the owner makes for a person who asks what is kept about them; the account page does not show it. It is deleted with the account.
Each account gets free credits once (since 2 November 2026), and every job uses credits (see Credits and payments). An account made before then also keeps how many free videos it started, how many more Takehound's owner gave it, and when its first and latest began. They are deleted with the account.
After you sign in with Google for the first time, the account page asks two questions, and you can skip either. The first is whether your video earns you money: from your channel, from a program or coaching you sell, for your business, or not yet. The second is where you heard about Takehound and, if an AI suggested it, what you asked. If you answer, Takehound keeps which choices you pressed, the words you typed if you typed any, and when, with your account (the first question since 8 November 2026, the second since 5 November 2026). Kyle reads the answers to learn who Takehound is for and how people find it. The owner's numbers count them, including by which post or ad brought the account, and list the words typed without saying whose they are. The answers are never sent to Google, Meta or any other ad or analytics service, and they are deleted with the account.
Since 8 November 2026, with your account Takehound keeps which of the three moments the editor tells Google and Meta about (see What the editor tells Google and Meta) have come for it, and when, so each is sent only once. An account made before then was marked as already signed up, and as having made a first cut where it had started a video, so nothing is sent for what it did before. It is deleted with the account, and is in the copy of an account the owner makes for a person who asks what is kept about them.
Kyle may use your account's email address to write to you about your account and about Takehound. Nothing in your account is shared or sold.
Takehound gets no access to anything in your Google account and never sees your Google password. The account page lists each Google account that can open your account. There you can remove one, or sign out everywhere.
Until October 14, 2026 an account could also be made with a passkey. Passkeys no longer sign in. An account made with one before then keeps what it had: Takehound still stores the public half of each of its passkeys and the name typed for it, if one was, and the account page lists those passkeys. A browser still signed in to such an account can add a Google account to it, and from then on that Google account opens it and the passkeys can be removed on the account page. The note your browser kept of which passkey was on your device is removed the next time you open the account page.
The account page loads Google's sign-in button from Google when it opens, so Google learns that your browser opened that page. Signing in sets one cookie of ours, so you stay signed in. Opening the account page sets a second cookie of ours, a short-lived one, used to check that a Google sign-in was started in your browser. Google's button keeps a small cookie of its own on this site, called g_state. On an iPhone or iPad a Google sign-in goes to Google's site and comes back, and for those few minutes there are two more cookies: one that Google's button sets, called g_csrf_token, and one of ours.
The list of email addresses
When the month's budget for AI models is spent (see The month's model budget), a Google sign-in that would make a new account is held, and goes no further. Takehound then keeps the email address on that Google account on the list below, with when it was kept and that it came from a held sign-in. Nothing is sent to it automatically. Kyle may write to it when new accounts open again, which is on the first day of the next month.
Until 2 November 2026, Open the editor took someone who was not signed in to a waitlist, and a Google sign-in from someone not let in put the email address on that Google account on the list, with no account made. Since then anyone can sign in with Google, and nobody joins the list that way. If you joined it, Takehound keeps your email address, the channel or site and what you make if you give them, what you say about a recording that never got edited if you mention one (it moves you up the line), the page you were on your way to, where your visit came from (a search engine, a site that linked here, or a campaign's name and which post or ad it was, the same thing the visit counts below record), the page you joined on, whose link brought you if a friend's did, a short code for your own link, when you joined and when you last joined again, and when Kyle let you in. Kyle may use the list to write to you about Takehound. Nothing on the list is shared or sold. It is kept in a file of its own on Takehound's server, apart from accounts, until you ask to come off it or Takehound deletes the list. To limit how many addresses one connection adds, joins are counted against the same one-way code as the daily limits below, never the IP address itself, and those counts are cleared each day. To come off the list, write to the address at the end of this page.
The try box on the home page
The home page can play a clip of yours without its pauses, ums and doubled words, with no account. Your browser reads the clip you pick and makes a copy of its sound only, small and in mono, and sends that to Takehound's server. The picture is never sent. A speech model (Whisper) that runs on Takehound's server turns the sound into words with their times; no other company receives it. The sound is deleted as soon as it has been read, and the words are held only until your browser collects them, and in any case for no more than an hour. Your browser then works out the cuts and plays your clip without them; nothing about the cuts is sent. If you press the sample instead, nothing of yours is sent at all. Takehound counts that a clip was tried, and how many each connection tries a day, against the same one-way code as the daily limits below, never the IP address itself; those counts are cleared each day.
The free shot list
The B-roll shot list generator at /tools/b-roll-shot-list (where /shot-list now leads) needs no account. The pasted text is sent to our AI provider to make the list, and it isn't stored. In detail: when you press Make my shot list, the script you pasted, and what you do if you filled that in, go through Takehound's server to a small text model, Google's Gemini, through OpenRouter, which relays it. Takehound asks OpenRouter to pass it only where what is sent is not kept (zero data retention) and not used to train a model; if that cannot be met the request is refused. Takehound's server keeps nothing of your script, of what you do or of the list: all of it is gone once the answer reaches your browser, and the boxes you tick, the copy and the download stay in your browser. What each list cost goes into the month's model budget as a number, and that a list was made is counted, by day, with nothing about who asked or what was asked. To limit how many lists one connection makes, they are counted against the same one-way code as the daily limits below, never the IP address itself: the last hour's count is held in the server's memory only, and the day's counts are cleared each day.
The same shot list can be asked for from inside an AI assistant, such as ChatGPT or Claude, once you add Takehound there as an app or connector, at the address the shot list page gives. Then the assistant sends the script you give it to Takehound's server, with the name the assistant's software calls itself by, and it goes on to the model and is kept nowhere, exactly as above. The list goes back to the assistant with a link to this site that names which assistant it came from, so we can count visits from each. That a list was made this way is counted by day with the name of the assistant it came from (ChatGPT, Claude, or another), and nothing else. What the assistant itself keeps of your conversation is covered by its own privacy policy. Lists asked for this way are counted against daily limits of their own, by the same one-way code.
The month's model budget
Takehound pays for each call it makes to an AI model. It writes down what each one cost, in dollars, added up by kind (the editor's agent, the first pass, the review, describing footage, the older tools) for each month, and how many sign-ins were held that month. These are numbers only: nothing about who asked or what was asked. When a month's total reaches Takehound's budget, new accounts wait for the next month (see The list of email addresses). Accounts that exist keep working.
Credits and payments
When you buy credits you pay on Stripe's own page. Your card details go to Stripe and never reach Takehound. Stripe tells us the amount, the payment's id and that it was paid, and it keeps your receipt and payment record under its own privacy policy. With a monthly plan your account also keeps the ids Stripe gives your subscription and its customer, the plan's status and when its month ends. Your account keeps your balance, what you have bought in total, and every move of your balance with why: credits given, bought or used, on which kind of job, and the price test or offer it belonged to. It goes when your account does, and you can read it on the account page.
So that free credits, an invite's credits included, are given once to a Google account, a one-way hash of the id Google gives you is kept after your account is deleted, with no address, no name and nothing else beside it.
Each account has an invite link with a code of its own, kept with the account. Opening someone's invite link sets a first-party cookie, th_ref, that holds that code and nothing else, for 30 days, or until a Google sign-in makes your account, when it is removed. The link then takes you to the home page with utm values saying an invite link brought you, and never with the code in the address, so the code is not sent to Google Analytics or the ad pixels a page may load. If that sign-in makes a new account, the account keeps which account's link it came by (that account's random id, never its name or address) and, once you export after your first video with Hound, what became of the reward: given to you both, or to you alone because the account that invited you had reached its limit or was deleted, and when. The account that invited you sees only how many friends joined by its link and how many rewards it earned, never who.
A link in an ad, or a members' link a creator community shares, can name an offer. A first-party cookie, th_offer, then keeps that offer's name for 30 days, so you are shown the same prices, and nothing else is in it. If you open an account, the free credits it starts with are kept with that offer's name.
Templates
Anyone can read the templates on the templates page with no account. Using one in the editor needs you signed in, like the rest of the editor, and so does publishing one. When you publish a template, Takehound keeps the template itself (its kind, its name, what you wrote about it, and the look, grade or steps it holds), the name you chose to show as its author, which account published it, when, how many times it has been used, and how many times it has been kept (a video or a timeline exported from a project that used it). The template and that name are shown to everyone, on the templates page and in the editor. The account it came from and your email address are not shown. You can delete a template you published, and deleting your account deletes every template it published. Takehound can hide any template.
Anyone can report a template. A report is kept for Takehound's owner: what it says, which template it is about and when it was made, and nothing about who made it. It is deleted with the template. To count each use of a template, and each time it is kept, once a day for each connection, and to limit how many reports one connection makes, Takehound counts them against the same one-way code as the daily limits below, never the IP address itself, and those counts are cleared each day. How many templates an account published today is counted the same way, under a code made from the account.
In the editor, a template you make and keep for yourself stays in your browser, and with your account if you have one, in the same list the editor keeps of how you cut. When you use someone else's template, your browser keeps a copy so it can draw your cards again, and asks Takehound for it again to see that it is still published; Takehound counts that use as above. When you export a video or a timeline from a project that used someone else's template, the editor tells Takehound which template, once for that project, and Takehound counts it as kept, as above: the templates page lists the community's templates by that count. Nothing about your project or your footage goes with it. Sharing a template from the editor sends the template and the author's name you chose, and nothing else.
Between Takehound and Anthropic
On the way to Anthropic's API, what you send may go through OpenRouter, a separate company that relays it. Takehound asks OpenRouter to pass it only to Anthropic's own servers, and only where what is sent is not used to train a model. What OpenRouter does with what it relays is in its privacy policy.
Hound's core
Hound, the agent in the editor at /editor, runs on a small core of his own. Since 14 October 2026 it makes Do the whole first pass on a story, and does what you ask in your own words, whenever Takehound's server can run it, unless the switch in the editor's settings is set to classic (or ?harness=classic is in the editor's address); a lesson's first pass is made the usual way. On it, asking Hound for something sends what any job sends (see The editor at /editor): the words of your recording, a description of your timeline (file names, times, text on screen), your notes, what you told him about your footage and the list of how you cut, and a few small stills of your video when he looks. It also sends the conversation you have had with him on that video, so he knows what you asked before (a new session, in the editor, starts it again), and the name, description and arguments of each of the editor's tools he has found for the job, the tools he made for you among them. They go through Takehound's server to Google's Gemini model through OpenRouter, which relays it, and Takehound asks OpenRouter to pass it only where what is sent is not kept (zero data retention) and not used to train a model. When he watches part of your cut, a small copy of that stretch, picture and sound at 480p, goes with his question the same way and is let go once Gemini has answered, and only the answer in words stays in the conversation. Takehound can instead give Gemini the copies to watch and listen to itself, held for the job, at the start of each job and whenever he watches again (a setting on its side, and a choice in the editor's address for testing); that was the usual way until October 28, 2026. Takehound can also run the core on Anthropic's model in place of Gemini, as it did until October 15, 2026 (a setting on its side, and a choice in the editor's address for testing). Then all of the above goes to Anthropic's model through OpenRouter in the same way, with zero data retention and not used to train a model. Anthropic's own servers are not offered that way, so the model answers from Amazon's servers (Amazon Bedrock), which run Anthropic's model. When he watches part of your cut then, a small copy of that stretch, picture and sound at 480p, goes through Takehound's server with his question to Google's Gemini model through OpenRouter, the same way the review's copy goes (see The review); the server holds that copy in memory until Gemini has answered, and never on disk. Where Takehound turns it on (a setting on its side, off for now, and a choice in the editor's address for testing), before he places one of your clips he also watches it: a small copy of a few seconds of each clip he is choosing between, picture and sound at 480p, made in your browser from your file (or from the copy kept in your footage library), goes the same way with the file's name, and is held the same way. Takehound's server holds those copies in memory for that job only, at most two hours, and never on disk. Your browser keeps the copies of your clips it made while the editor is open, so a clip watched again is not made again; they are gone when you close the tab. The changes he makes to your cut are short programs he writes, which run in your browser, walled off from the internet, from this site and from what your browser stores. When he writes down a way of working for you, or makes you a tool of your own from the editor's tools, it is kept the way the commands the editor learns are kept: in your browser, and with your account if you have one. If the model cannot be reached that way, nothing is sent: a story's first pass is then made the usual way (see The editor at /editor), and any other job stops.
That conversation, Hound's session on the video, is kept in your browser and, if you have an account, with it, so a reload or another computer picks it up where it was: what you and Hound said, the tools he found, his plan, what you typed while he worked and where each of your asks began. The pictures he looked at are left out of what is kept. Takehound's server keeps it with the project, in Cloudflare R2 while footage storage is on, otherwise in its own database. Deleting the project deletes it, and Delete this account removes all of them.
The decision model
When Hound's core cannot make a story's first pass (see Hound's core), the pass is decided by Jev, a decision model made by TypeSafe AI, a separate company. It answers questions about text with a number and writes nothing. What it is sent is text only: the words of your recording, the descriptions of your clips and what you told the editor about your footage. It goes from Takehound's server through OpenRouter, which relays it. Takehound asks OpenRouter to pass it only to TypeSafe's own servers, and only where what is sent is not kept or used to train a model; if that cannot be met the request is refused and the editor does the job the other way. Hound also asks Jev quick questions while he works on anything you ask (does this clip fit this line, which opening is strongest). Each goes the same way, and is the text he writes for it: lines of your recording, descriptions of your clips and their file names when he puts them in. What TypeSafe does with what it is sent is in its legal pages.
References from the web
When you name a specific thing on camera, such as a book, a site or a product, Hound can look it up and show it beside you. Words from your recording go out for this, and only those words: the name, as you said it, with the author of a book or the address of a site when he knows them. They go from our server to Open Library, or to Wikipedia, Wikidata and Wikimedia Commons, or to the site itself, and those services see our server, not you. They are never sent your recording, your other words, a file name or a time, and a name that is not in your transcript is not looked up. These are public catalogues with their own policies, not model providers, so the zero retention above does not cover them. Our server fetches the picture you pick, removes its metadata and passes it to your browser, where it joins your footage library marked as a reference, with its source, and is kept and deleted like your other clips. Each one's page is linked beside it, and one press takes it off the video. A reference is not licensed by Takehound: it is yours to decide whether to publish it.
The review
After a first pass, and when you ask the editor to check your cut, the editor watches the cut back and tells you what is rough in it. When the editor reviews a cut, a small copy of the cut, picture and sound, goes through Takehound's server to Gemini, a video model made by Google, through OpenRouter, which relays it. Your browser makes that copy from the video as it stands and sends it in pieces of a few minutes, each with the times of the cuts in it and the words heard either side of each cut; no file names are sent. In the same review the editor also checks each of your clips it put on screen: a few seconds of the cut around the clip, picture and sound, goes the same way, with the words you say over it, the sentence before and after them and what you told the editor about your footage, and Gemini says whether the clip shows what you are talking about. Takehound asks OpenRouter to pass it only where what is sent is not kept (zero data retention) and not used to train a model; if that cannot be met the request is refused and the cut is not reviewed. Takehound's server keeps nothing of it once the review has answered. What OpenRouter does with what it relays is in its privacy policy. You can stop a review while it runs, and turn off the review after a first pass in the Takehound panel.
Training
Nothing you send is used to train a model. Takehound doesn't train on it, and Anthropic doesn't train on data sent through its API, under its commercial terms. What goes to Google's Gemini model is sent only where it is not kept and not used to train a model (see The review). What Hound's core sends to Anthropic's model through Amazon Bedrock is sent the same way (see Hound's core). TypeSafe says Jev is not trained on what is sent to it, on its models page.
The older tool at /app: your footage
In the tool, your video files are not uploaded. To describe a shot, your browser makes a few small still images. It sends only those stills to Takehound's server, which passes them to Anthropic's API to write the description. Takehound doesn't keep the stills: they are discarded once described. Your catalog of shot descriptions, with a small thumbnail of each shot, is stored in your own browser, so the same footage isn't described twice. The Clear my data button in the tool deletes it.
The older tool: your recording
If you add your talking-head recording, your browser uploads its sound to Takehound's server, which writes a transcript with timings and sends it back. The sound is transcribed on Takehound's own server and isn't sent to any other company. The sound file is deleted as soon as the transcript is made, and within one hour at the latest, even if something fails.
The older tool: your script and cut
When you make a cut, your script or transcript, the shot descriptions and the names of your clips and folders go to the server and to Anthropic's API to choose the shots. When you download the timeline, the server also gets the full folder path you typed (on a Mac it usually includes your user name), so it can write it into the timeline. Takehound's server holds the transcript and the cut in memory for at most an hour, so the tool page can pick them up, and keeps nothing of them after that: they live in your browser tab. The one exception is a cut you open in the studio (next). Keep client names out of your file and folder names.
The studio
If you open a cut in the studio, Takehound's server keeps it as a project so you can come back to it: your transcript, the cut and every change made to it, your notes, and the descriptions and file names of your shots. It is kept until you delete it. Anyone who has the project's link can open it, so share the link only with people you want to see it. Delete this project, on the project's own page, removes all of it.
The studio's editor
When you ask the studio's editor for something, or leave it a note, your transcript, the cut, your notes and your shot descriptions go to Anthropic's API so it can do the work. It works only on your project.
Daily limits
The tool has daily limits, counted per connection. Takehound's server turns your IP address into a one-way code, using a secret that changes whenever the server restarts, and counts uses against that code for the current day (UTC). It never stores the IP address itself, and the counts are cleared each day. The tool's steps (shots described, minutes transcribed, cuts made, timelines downloaded, cuts opened in the studio, jobs given to the editor, finished videos asked for, accounts made) are also counted under the first part of that code, with the visit counts below. When you make an account or open a cut in the studio, that step is counted at the moment the account or project is saved, so the steps counted under that code can be matched to that account or project.
Invite requests sent before October 7, 2026
Apart from the list of email addresses (see The list of email addresses), the only email address Takehound takes in is the one on a Google account you sign in with (see Accounts). Invite requests sent before October 7, 2026 (the email address and answers typed into the form, and which link brought that visit here) stay in a file on Takehound's server until Kyle deletes them. They are never shared or sold.
How we count visits
Takehound counts page visits and steps (for example, "played the demo" or "clicked Try it") with its own code, which doesn't set cookies or store your IP address. A random id lives in your browser tab until you close it, so the steps of one visit can be counted together. We also note which link brought you here (for example, a Reddit post) so we know which places are worth our time. If you then sign in with Google and that makes a new account, the link's utm values, the sending site's host name and the path of the page you landed on are kept with the account (see Accounts). Since 9 November 2026 some of our links are short ones, such as the one on our X profile: following one takes you to a page with that link's utm values added, and the server counts that the short link was followed, with the day and time and whether the browser was a phone's. That count has no id, no cookie, no IP address and nothing else about you, and it is made even if you leave before the page loads. Since 17 November 2026 the server also counts when an AI assistant's own fetcher reads one of our pages, for example when someone asks ChatGPT, Perplexity or Claude a question and it opens the page to answer: the fetcher's name as it gives it, the page and the day. Such a read is not counted as a visit, and the count holds nothing about the person who asked. Since 9 November 2026 the server also counts how each sign-in with Google ended (for example a new account, an account opened again, or a sign-in Google did not confirm) and which of its ways it came by, with the day and time and, since 10 November 2026, whether the browser was a phone's. That count has no id, no cookie, no IP address, no account and no email address. The server also counts three steps of invite links: that an account's invite link was opened, that a Google sign-in made a new account that came by one, and, when such an account exports after its first video with Hound, what became of the reward (given to both accounts, or to the new account alone because the account that invited it had reached its limit or was deleted). Each of these counts is kept with the day only, never the time, and has no invite code, no id, no cookie, no IP address, no account and no email address. Since 10 November 2026 the account page counts a press on Google's sign-in button as a step, and, for someone signed in on a phone, that the card saying the editor needs a computer was shown and which of its buttons was pressed. The link on that card is a short one, and it leaves your phone by the phone's own share or mail app: Takehound sends no email and does not learn where you sent it. In the editor, the same counting notes a few first steps: opening the sample project, leaving it to start your own video, and asking Hound something. Since 9 November 2026 it also notes, after the first export in a project, that the editor showed the card for your next video, and whether you pressed its Start your next video. Since 10 November 2026 it also notes, on a phone or a tablet, that the editor showed its screen saying it is made for a computer, and whether you pressed Email yourself the link or Take a look anyway; the link in that email carries utm values naming that screen, so a visit from it on your computer is counted as coming from it. Since 15 November 2026 it also notes when your browser would not keep a file you added, and whether that was because its storage was full or for another reason (a private window can do this); the file's name, size and contents are not sent. Since 16 November 2026 it also notes when a tab opens the editor from our Shorts tool page, and when the Shorts job that page is for starts in it; nothing about the video is sent. When you are signed in, each of these steps carries a tag made from your account, never your name or address, so we can tell how many people opened the sample and then asked Hound about their own video. Takehound's owner can also see, in the list of accounts with their email addresses, which link each account's first steps in the editor came from (since 8 November 2026). Nothing new is kept for it: the list reads the tag and the link from these counts. An invite request sent before October 7, 2026 was joined to the visit it came from. The pages our ads link to, whose addresses start with /start/, ask what you pay an editor for one video; Takehound's own counting records the answer you press, which is not sent to Google or the ad companies.
Fonts
Every page, this one included, loads its fonts from Google Fonts, a Google service. When a page opens, your browser can fetch the fonts from Google, which then sees your IP address and browser details. This happens on every page, including the ones that load neither Google Analytics nor the Meta Pixel.
Video from YouTube
Three guides show Kyle's Episode 1 from YouTube as a picture with a play button. Nothing is fetched from YouTube until you press play. Then the video loads in YouTube's privacy-enhanced player (youtube-nocookie.com), and Google sees your IP address and browser details, as it would on YouTube.
Badges
The home page's footer shows seven small badges from sites that list Takehound: Fazier, Startup Fame, Findly, Good AI Tools, Aura++, ScrollLaunch and Noonlaunch. Your browser loads each badge from that site, which then sees your IP address and browser details.
Google Analytics
The landing pages (the home page, /coaches, and the pages our ads link to, whose addresses start with /start/) also use Google Analytics, a Google service, to see how visitors find and use the site. It sets its own cookies and tells Google that you visited and which steps you took. It isn't loaded on the tool page (/app), and it isn't loaded when your browser's time zone is in Europe.
The Meta Pixel
The landing pages (the home page, /coaches, and the pages our ads link to, whose addresses start with /start/) also load the Meta Pixel, a service from Meta, so that ads we run with Meta can be measured. It sets its own cookies and tells Meta that you visited the page. Meta sees your IP address, your browser's details and the page's address, as it does on any site that uses the pixel. Its automatic events are switched off, so it does not report the buttons you press. Like Google Analytics, it isn't loaded on the tool page (/app), and it isn't loaded when your browser's time zone is in Europe. The Meta Pixel also hears the editor's three moments below. No other ad company's pixel is on.
What the editor tells Google and Meta
When a new account opens the editor, when its first cut is finished (a first pass, or an ask to Hound that changed the cut) and when it first exports a video or a timeline file, the editor tells Google Analytics and the Meta Pixel. This lets an ad be measured by the people who signed up and made something, not only by clicks. Each is sent once per account. What is sent is the event's name (sign_up, first_cut or export; Meta is told a sign-up under its own name for one, CompleteRegistration) and where your visit came from (for example, reddit), and nothing that names you: not your email address, your name or your account. Neither is loaded in the editor until one of these is due, and neither is loaded or sent when your browser's time zone is in Europe. When they load, they set their own cookies, and Google and Meta see your IP address, your browser's details and the address of the editor's page, as they do on any site that uses them. Google is given the page's path only, and Google's advertising storage stays off. Separately, the editor tells Takehound itself, and no one else, when your account exports (once each time the editor is opened), also where Google and Meta are not loaded: it is when an invited friend's credits are given. Nothing about the video or timeline is sent with it. Your account keeps the date and time of that first export.
Do Not Track
We don't respond to Do Not Track signals. Our own visit counting works the same either way; Google Analytics and the Meta Pixel follow their own rules.
Changes
If this page changes, we update it here and change the date at the top.
Questions or deletion
To delete the shot catalog your browser keeps, press Clear my data in the tool. A studio project is deleted from its own page, a clip of your footage library with Remove in the editor's library, and an account from the account page. For a question, or to ask that an account or anything else Takehound keeps about you be deleted (an invite request and its code's step record included), write to kyletran77@gmail.com. Takehound keeps a daily copy of its accounts file for three days, a week at the most, so a deleted account is gone from those copies within a week. Our host's backups, and its logs of invite requests sent on or before October 3, 2026, may keep a copy until the host deletes them on its own schedule.