What Commentware stores

Commentware takes screenshots of web pages and keeps them, so it matters exactly what is stored, where it is, and what moves. What the people who run the hub hold is in the Privacy Policy, and the licenses are on the License page.

The server is yours. It runs on a machine you choose, and what it stores never reaches the hub, so you decide what happens to it — and when you run a server for other people, you answer to them for it.

What is stored

All of it is on the machine running commentware start — the SQLite file and the directory beside it.

The two in bold are the ones you did not choose field by field, and they are the reason the switch below exists.

When the element sits inside a component the page keeps hidden from other scripts, the same kind of text is kept for the component around it as well, and the switch below stops both.

When the element sits inside a page embedded in the page you comment on, by an iframe or an <object> tag, and the embedded page is from the same site, the same kind of text is kept for the element in it, and the switch stops it. When the embedded page is from another site, the ways to find the element again (its path, id and role) are kept, with the start of the embedded page’s address without any query, a fingerprint of the element’s text and its length, and never the first 128 characters of that text, whatever the switch says.

It stays until you delete it. Uninstall Commentware asks whether to delete it, and keeps it unless you type y. A y deletes your account’s Commentware data directory, with the database, every screenshot, drawing and attached file, and the logs in it, and the directory that holds the server’s own config file when that is a different one. A database your config file keeps somewhere else is named and kept, and an export file you saved is yours to delete.

A directory of screenshots, drawings and attached files that your config file keeps somewhere else is named and kept too.

An export file an AI agent uploads to be imported is held beside the database, in a directory named after it, such as data.db.imports, while the import runs, and is deleted when the import ends, whatever the outcome. If the server stops in the middle, the next commentware start deletes it. The directory stays, so it goes with the data directory when you uninstall, and stays beside a database your config file keeps somewhere else. Learn more about importing threads through an agent.

A screenshot is the largest thing here. A capture of a viewport can contain a session token in a devtools panel, a customer record, an inbox — much more of the page than the comment was about.

Capture can be switched off

Screenshots can be switched off for one site or for the whole server. Off refuses new captures and new drawings, and leaves stored ones readable. A file attached to a comment is not a capture, so the switch has nothing to say about it. The short piece of page text kept with an anchor goes with it, and a page title does not. Learn more about what is kept about the page.

What moves, and what does not

• The server makes no outbound call that leaves your machine. No telemetry, no update check, nothing. The one request it ever makes is to itself: when an agent authenticates in a browser, the gate reads the key that verifies its token from the server’s own address. That dials the socket this process opened, and it only happens at all on a server whose address is loopback or HTTPS.
• The extension makes one request of its own, to the package registry. Once a day it asks the registry, at https://registry.commentware.com, which version of the extension is newest, so it can tell you when yours is behind. The Privacy Policy says what that request carries and what the registry learns from it. That is separate from the first bullet: the server still makes no outbound call, and has no update check.
• A screenshot, a drawing or an attached file leaves the server only to somebody holding a key. The route that serves them, and the agent tools that read them, are behind the same access key or agent token as everything else; there is no unauthenticated read. What happens next is the key holder’s: the extension shows it in your browser, an agent that reads one passes it to the model it runs on, and the options page’s Export / Import screen, which is experimental, writes the screenshots, drawings and attached files of the threads you export into a file you choose. That file carries no key: anybody who has it can unzip it and open each screenshot, drawing and attached file. Nothing sends the file anywhere unless you import it, which uploads its screenshots, drawings and attached files to the server the extension is connected to. See Export and import threads.
• Your browser only asks a server about a page on its own once that server has told it there are already comments on that site. The icon’s badge count and the side panel’s list are both held to this; opening the comment overlay is different, because that is something you did, and it always asks, known site or not. The extension has permission to run on every page; nothing that asks on its own asks about one nobody has commented on yet.
• The extension uses your clipboard only when you do. It writes to it when you press a copy button — Chrome lists this as “Modify data you copy and paste” — and reads only what you paste into a comment. The website you are on cannot see the copy happen, so it cannot catch what is copied or swap it on the way. Like any website, it can still put something of its own on the clipboard shortly after you click anywhere on it, the copy button included.
• One thing can leave, it is Chrome’s, and you choose. Dictated audio may be recognized by Google’s speech servers, because that is what Chrome’s speech API is. Nobody is defaulted into it, and “only on this computer” is enforced by Chrome rather than promised by Commentware. No audio ever reaches a Commentware server. See where your voice is recognized. By default the microphone is open while an overlay is, and for five minutes after, so that the first word is heard; nothing is recognized or sent until you press the button. See when the microphone is open.

What the installer keeps and asks for

The installer, commentware-installer-cli, is a program of its own, apart from the server and the extension. It keeps your invite code in ~/Commentware/invite (%USERPROFILE%\Commentware\invite on Windows), with mode 0600 on macOS and Linux. Deleting the file removes the code, and so does Uninstall Commentware. Learn more about the invite code.

On every run it asks two hosts. It asks the package registry, at https://registry.commentware.com, for each part it installs or updates — the extension, the server, or both, as what a run installs and updates describes. It asks the hub for the list of invite-code digests for the batch your code names, at /codes/<batch>.json. The address of that list holds the first four characters of your code and nothing else about it.

The installer asks nodejs.org for a file only when it downloads Node.js: when it sets up a server and finds no Node.js to run it on (it asks first when it has a terminal to ask on), or when it moves its own copy of Node.js to a newer version on an update. It sends nodejs.org no code and no key; nodejs.org sees the address, the time, and a User-Agent naming the installer and its version. It asks no other host.

What is not defended

Stated plainly, because a security claim that is not true is worse than none.

• Nothing is encrypted at rest. The database is a file, every screenshot, drawing and attached file is a file, and the browser’s copy of your access key is on disk. Anybody with your machine has all of it. This is a program you run on your own computer; defending it from its own operating system’s owner would mean a passphrase on every command for a threat that already has the machine.
• Plaintext over the network. Bound to anything but loopback with no TLS in front, everything this server accepts crosses the network in the clear. The server warns you when you do it.
• A can_collaborate key can write a lot. It cannot delete, but it can resolve every thread on a server and comment in all of them. Everything it does is reversible, which is the property that makes the grant safe to hand out; it is not a quota.
• A server sees every comment written to it. That is what a server is for. If you connect to somebody else’s, they can read what you write to it.
• What is filtered about a page’s address is its shape, never the page. A small set of query parameters that look like a credential are withheld before the address is sent: when the extension asks what is on a page, and when you place a comment, move a page or import an export file. A parameter named token, jwt or session, a value shaped like a JWT, and every parameter on an OAuth or SSO callback path are withheld. Nothing here reads what is on the page to decide what is safe to send, and a URL shaped in a way this does not recognize still travels. A Chrome profile dedicated to the sites you comment on is the backstop for what a URL alone cannot show — see a Chrome profile only for this.

Commentware is for sites you are building

Your own domain, your own staging, your own localhost. On a page you do not control, a hostile site has real reach: it can read what you type into the composer, and it can hide or move the comment overlay entirely.

Some of that is defended and some of it is not, and the line is drawn where the defense is cheap rather than where it is absolute. It is written down so it can be disagreed with, and revisited the day people start commenting on pages they do not control.

Two things it cannot do, which are the ones that would matter most: it cannot reach your access key, and it cannot make Commentware act — every control that ends in a write needs a real gesture from a real person, and a click a page dispatched at one is refused.

Comments are never treated as markup

A comment body is text a stranger wrote, drawn on somebody else’s page. It is parsed and drawn as elements, never turned into HTML, so no comment can contain anything that runs. Links are clickable only for http, https and mailto; anything else stays readable text.

Next steps

• Settings: switch capture off for a site or for the whole server.