You already did it once. That should be enough.

Because the voice is in the box. GuideGen narrates your video with a neural voice that runs on your computer — the model and its pronunciation dictionary ship inside the extension, which is about 60 MB of it. Tools that download in three megabytes do the narration on their servers, which means the script of your internal process is sent to them. That is the whole trade, and it is the one we would make again: you pay for it once, at install, and then nothing you record ever leaves your machine. A one-off 67 MB is a few seconds on any connection made this decade. Moving the voice to a server would make the download smaller and the promise worthless, so we are not going to.

Nothing, until you press Publish on one specific guide. Recording, editing, redacting, and every export — video, PDF, slides, Markdown, the AI handoff — happen entirely in your browser. When you do publish, that one guide's steps and its already-baked screenshots go up so the link works, and nothing else does: not your other guides, not the API log, not anything you redacted, because redaction is burned into the image before it is ever sent. The privacy page names every field.

Guides live in your browser's local storage, so removing the extension removes them. Export anything you want to keep first — the HTML, PDF and Markdown exports are self-contained files with the images inside them, and a published link keeps working because that guide is already hosted. There is no lock-in to escape: every format is one you can open without us.

Really free, unlimited guides, no watermark, no trial, no card. The honest reason: GuideGen is part of backpocket.website, which is a portfolio — it exists to be good, not to be sold. Nothing costs us per guide, because the recording, the narration and every export run on your machine rather than ours. The only thing we host is a published link, and that is a small enough bill to keep free. There is no paid tier waiting to take features away.

No, and the difference is worth thirty seconds. Backpocket AI is ours: it is the voice that reads your narrated video, and it runs on your own computer — the model ships inside the extension and nothing is sent anywhere to produce it. The AI handoff is the opposite direction: it puts the written workflow on your clipboard so you can paste it into whichever assistant you already use. Nothing is sent there either until you paste it yourself. Backpocket AI, defined in one place — where it runs depends on the tool, and each page says which.

All of them. It copies plain text to your clipboard — Markdown, with the steps grouped under the page each happened on — so it pastes into Claude, ChatGPT, Gemini, Copilot, an IDE, a Jira ticket or a plain note. There is no integration to configure and no API key to hand over, which also means nothing is sent anywhere until you paste it yourself.

No, and that's deliberate. Forty screenshots as embedded data would be several megabytes of text that no chat window will accept, and a model told to expect images that never arrived will describe them anyway. The handoff is the written record — the steps, the actions and the URLs. When the pictures matter, attach them from the Markdown or PDF export, or send the shareable link.

No, and that was the design problem. A single-page app fires requests on nearly every click, so printing them all under every step would bury the guide. Instead a step where everything succeeded shows one muted line — 5 requests — that doesn't expand. A step where something failed shows the failures straight away, because that's the case you opened the guide for. Everything, in full and untrimmed, lives in one API log panel you open from the toolbar, with a "failed only" filter and a copy button.

Only the AI handoff. HTML, Markdown, PDF, PowerPoint, narrated video and published shareable links all leave it out — a guide someone is following doesn't want GET /api/session → 200, and a video certainly doesn't. If you want just the log and not the guide around it, there's a copy button inside the log panel.

No. A header whose name looks like a credential — authorization, cookie, x-api-key and the like — keeps its name and loses its value, replaced with a marker before anything is stored. There is no setting that changes that. The same happens to query-string values and to obvious secrets in a submitted body, so a password field never lands in a log. Your browser cookies aren't readable this way at all. The practical consequence: the cURL documents the call, it does not replay it — paste your own token in if you want to re-run it.

The full exchange is itself an opt-in that's off by default, and worth being deliberate about: with it on, successful responses are kept too, and those often hold real data — customer records, a whole result set. Turn it on for the recording where you're capturing a bug, not as a standing default. Bodies are trimmed, credential-looking keys are masked even inside them, any step's log can be deleted from the editor, and none of it is ever uploaded — publishing a guide never publishes its log.

No. It records while you have a recording running, and — separately — on any single site you have switched catch-up capture on for. Nothing else. Catch-up is off everywhere until you turn it on for a specific site, one site at a time, and while it is on there is a dot in the corner of the page saying so. What it holds stays on your device, expires after seven days, and is never uploaded; you can also discard it yourself at any point from your dashboard.

Because on a site you've armed it has been capturing all along and discarding as it goes, keeping only a rolling window. There is no clever trick and no way around it — screenshots of the past do not exist, so something has to have been running. That is exactly why this is per-site and off by default rather than always on everywhere, and why the page shows a dot while it is happening.

No. GuideGen records what a control was called and where it was on screen, not a selector that would survive the page changing — so the handoff is for explaining and reasoning about a flow, not replaying it. If you want a script, a browser automation recorder is the right tool; this one is for when a person or a model needs to understand what happened.

No. GuideGen documents web applications — anything you can open in a Chromium browser. It cannot capture native desktop software.

Yes. Press Publish on a guide and you get a link anyone can open. Only the guide you published is uploaded, and its annotations and redactions are burned into the images before they leave your machine, so no unredacted original is ever sent. You can update it in place — the link doesn't change — or unpublish it, which kills the link and deletes the images. If you'd rather not publish at all, the self-contained HTML export is a single file you can drop into Google Drive, Notion, SharePoint or an email.

Two reasons. The editor lives on this site rather than inside the extension, so there is one editor to keep good instead of two, and it needs to know whose guides to show you. And publishing has to be tied to an identity so that you — and only you — can update or delete a shared guide later. Signing in does not upload anything: guides you have not published stay in your browser's own storage on that machine.

Published guides, yes — sign in anywhere and they're there. Unpublished ones, no: they live in the extension's storage on the machine that recorded them, which is the direct consequence of not uploading anything until you ask. That is the design working, not a limitation we forgot about.

Step descriptions are generated in English at the moment. It will happily record an interface in any language, but the sentences it writes around it are English.

Because it cannot know in advance which application you'll want to document — it might be any internal dashboard. The recorder does nothing until you press start — or, on a site you've switched catch-up capture on for, until you work on that site — and it reads only the label and position of the element you click, plus the request summary described above. It also asks to observe network requests, which is what the API log is; it cannot block, redirect or alter one, and credential values are masked before anything is stored.

Yes, and with no card required. Paid plans may appear later for team features; if that happens, people already using it get notice and a choice rather than a surprise bill.

About 3 MB. The neural voice that makes narration work offline is another 78 MB, and it downloads the first time you export a narrated video — once, on that machine, with a progress bar. Everything else is ready the moment it's installed, and the voice is only fetched if you use the feature that needs it. Your text still never goes to a server: the voice comes down to you, your words never go up.