What the TweakThis extension captures

When it runs

The extension works only on sites you added to a project, matched by the project’s URL rules (scheme, host, port and path; a wildcard host covers subdomains). You grant it each site once. On every other page it does nothing, and your browsing history is never read or sent.

What you send

• What you point at: one element, a box you drew around an area, or the whole page.
• Your note (up to 4,000 characters), and if you like what you expected and what happened instead.
• A category and a priority, if you change them.
• Any arrows you drew and areas you blacked out on the screenshot.

The page

• The address and title. Before anything is sent, user names and passwords in the address are removed and values of parameters that look like secrets are replaced (see how redaction works).
• The viewport and page size, scroll position, zoom and device pixel ratio, the browser and its version, and the user agent.

The element and its context

• Up to 8 selector candidates, built from ids, test ids, labelling attributes, nearby headings, roles and text, with a CSS path as the fallback. Each says how many elements it matches, and the ones that match only this element come first.
• An excerpt of the element’s HTML with its ancestors, the text of its siblings and its labels, 16 KiB at most. Script and style bodies, comments and value attributes are left out.
• Up to 20 attributes (id, class, role, href, aria attributes, test ids and the like), never value.
• Key computed styles: display and position, size, margin, padding and gap, flex and grid settings, font, line height, colour and background, border, radius, shadow, opacity, z-index and overflow.
• Its visible text, accessible name, any text you selected, its position on the page, and its path through shadow roots when it is inside one.
• For a box you drew: the visible text inside it and the element at its centre.

Screenshots

One screenshot of the visible part of the tab, not the whole page, and a crop of the element or box from it. Fields that hold secrets are covered while it is taken. Areas you black out are removed from the pixels in your browser before upload, so the original never leaves it. The agent gets a preview sized for image input (at most 1568px on the longest edge) and a link to the full image.

Guided flows

To show how to get somewhere, you can record a guided flow instead of a single note. For each step it records the page address and title, the element you clicked and its text, scrolls and page changes, and the label and type of form fields you changed, never what you typed in them, with your notes and screenshots of up to 10 steps.

What it never reads

• Values of input, textarea and select fields: what anyone typed into a form. Text in an editable area of the page is page text and is captured like any other.
• Console logs or network requests.
• The inside of frames from another site: an iframe is recorded only by its address, with a note that its content could not be read.
• Pages that are not covered by one of your projects.

Where it goes

Send shares a note right away. With a review session open, notes stay on your computer until you publish the session, and nobody sees them before. Once shared, the project’s team sees them in the dashboard, and the coding agent you connect reads them over MCP: see the setup for Claude Code or Codex. How long screenshots are kept depends on the plan; the privacy policy has the details.