The steps of a flow as you perform it: a screenshot at each step, the element you interacted with, the page URL, and the text of what you clicked. It is not a continuous recording of your desktop, it does not run in the background, and it covers only the tab you started it on. Capture begins when you press record and ends when you press stop.

No. It is inert until you start a capture and it acts only on the tab you started it in. There is no background collector, it does not read your other tabs, and it has no access to anything you have not deliberately pointed it at.

Uploads carry a capture token scoped to a single flow, not your account session. A token that leaked could write to the one flow it was issued for and nothing else: it cannot read your workspace, reach another organisation, or act as you anywhere in the product.

Delete the step or the whole flow, and the screenshots, video and GIFs belonging to it are removed with it rather than left orphaned in storage. We do not mask sensitive fields automatically today, and we would rather tell you that than let it be assumed, so the honest advice is the one we follow ourselves: record against staging or a seeded account when a flow touches real customer data.

Postgres on Neon for the structured data, Cloudflare R2 for screenshots, video and GIFs, and Vercel for the application itself. Choice of data region is part of a Business plan and is agreed on the call that sets it up, rather than being a switch you toggle yourself.

TLS on every request, and AES-256 at rest for both the database and your stored recordings. Nothing leaves your browser in the clear and nothing sits on disk unencrypted.

Projects, flows and recordings belong to the organisation that created them, and every read and write is checked against the organisation on your session before the query runs. To be precise about the mechanism, because a reviewer will ask: that check is enforced in the application, not by row-level security in the database. The database is reachable only by the application's own role, so there is no second path to the data, but we are not going to describe an application control as a database one.

Not another organisation's. Every read and write goes through our servers, so the browser never holds a database key. There is no client-side credential to steal that would widen access beyond the account already signed in.

Production access is limited to the engineers who need it to operate the service, and we do not browse customer content as a matter of course. We will not claim an approval workflow we have not built: today this is a small team and a short list. If your assessment needs named individuals and a documented process, ask on the call and you will get both in writing.

Not on Business, where it is part of the plan: captures, recordings and documents from a Business workspace are never used to train any AI model, ours or a provider's. On Free and Small scale we reserve the right to use workspace content to improve our own models, and the comparison table on the pricing page says so against the plan. We run on enterprise API plans either way, so no third-party provider trains on your content on any tier.

The captured steps and their screenshots go to the model provider over TLS for the length of that request, and the finished document comes back. The provider is under an enterprise agreement with training disabled. Nothing is sent to a model you have not triggered: capturing a flow does not call one, and a document is written only when somebody asks for it.

On Business, yes. Inference then runs on your provider account, under your agreement and your retention terms, and never touches our model credentials. It is the strongest answer we can give a team whose policy is that inference stays inside their own vendor relationships.

For as long as the account is open, because the product is a library of your flows and a retention clock that deleted them would delete the thing you bought. Custom retention and deletion windows are part of a Business plan. Delete a flow, a project or a document at any time and its stored media goes with it.

Ask and we delete it, across the database and object storage, and confirm in writing when it is done. Backups age out on their own cycle rather than being surgically edited, which is the honest answer every provider should give and most phrase more vaguely. Anything already published to your own docs site over MCP is in your system and stays there.

Vercel for application hosting, Neon for Postgres, Cloudflare R2 for stored media, Firebase for authentication, and our AI model providers for the inference a document or a QA check needs. You get the current list with the DPA, and we tell you before we add one.

No. Nothing is shared until somebody deliberately shares it, and sharing is per document rather than per workspace, so publishing one page does not expose the flow it came from or anything else in the project. Viewers see only what they have been given.

Authentication runs on Firebase, so we never hold a password in any form we could read, and sessions are verified server-side on every request rather than trusted from the browser. Single sign-on with SAML and SCIM provisioning are part of a Business plan and are set up on the call.

An audit log of captures, exports and shares is part of a Business plan. On the other tiers there is no customer-facing audit trail today, and we would rather you knew that before you asked for one.

Vercel and Neon, both SOC 2 Type II certified. To be clear about what that covers: the infrastructure is certified, Zorva itself is not yet, and we would rather tell you that up front than let it be assumed.

SOC 2 and ISO 27001 for Zorva as a company, row-level security in the database, and automatic redaction of sensitive fields during capture. Those are the four a thorough assessment will find, so here they are first. We will tell you what is in progress and what is not started, and we will not describe a roadmap item as a control.

Email security@zorva.ai and a person answers. We will not threaten anyone who reports something in good faith, we will tell you what we found and when it is fixed, and we will credit you if you want to be credited.

Please do, and before you put real product footage into it. We will complete your questionnaire, sign a DPA, walk your team through how the whole thing works including the parts we are still hardening, and answer whatever this page did not think to ask. If something in your assessment is a hard requirement we do not meet, we would rather hear it now than ship around it later.