Tools

install

Sets up Page Scanner's helper for every Chromium browser on the computer, and returns which browsers and what the user does next: press Connect in the extension's settings and allow Chrome's prompt. Returns no secret. An unpacked build loaded in developer mode is found and allowed by itself. One optional argument, extensionIds: more extension ids to allow by hand. See Connecting Chrome.

pair

The older setup. Writes the pairing and returns the port and token to paste into Chrome. One optional argument, rotate, which issues a new token and invalidates the old one. See the note in Connecting Chrome about the token reaching the transcript.

list_browsers

No arguments. The Chrome profiles connected right now, with the name you gave each one and how long it has been connected. Start here when more than one is connected.

list_tabs

The windows and open tabs of one browser. Tabs carry the windowId they belong to, and windows say which is focused, so an agent can pick a window as well as a tab.

scan_page

Captures a page, or a list of them, and writes each to disk. Returns the absolute path.

A tall page or a long list can take the browser minutes, longer than some clients wait for one call (60 seconds is the MCP SDK's default, and Claude Desktop's). A client that asks for progress gets a notification every 10 seconds and the result at the end. One that does not gets status: "running" after 45 seconds, with startedSecondsAgo and, for urls, pagesDone and pages, while the capture goes on in the browser. Call scan_page again with exactly the same arguments to wait for it: the call joins the running capture and returns its result, which is kept for 10 minutes if it finishes while nobody is waiting. A capture of the same tab with other arguments is refused until the first ends, because only one debugger can drive a tab at a time. extract_design works the same way.

What comes back

The absolute path, the capture's width and height in CSS pixels, the mode (vector or raster), selectableText, and truncated.

selectableText is true only for a PDF that came out of the vector pipeline, which is the thing an agent should check before promising the text is real. truncated is null when the capture was whole, and otherwise carries what the page measured, what was captured, and a sentence naming the gap.

When anything was hidden, by hide or by the extension's settings, the result has hidden: how many ads, consent banners, chat widgets and overlays the scan took out of the page before capturing it.

When the page declares how to cite it (citation tags, schema.org data, Dublin Core, OpenGraph, a canonical link), the result has citation: kind (journal, conference, thesis, report, book, news, blog or webpage), and those of title, authors, published, modified, container (the journal or site), publisher, volume, issue, pages, doi, isbn, issn, canonical and language it gave, as it wrote them and unchecked, with from, the vocabularies they came from. The file carries the same. A page that declares nothing, and an extension older than this, give no citation.

When the browser had machine-translated the page before the scan (Chrome's or Edge's translator), the result has translated: { "by": "chrome" } or "edge". The text is then the translation, not the page's own words; the Markdown's front matter says translated: and the PDF's Subject says so too. Quote such a page as a translation.

With markdown, the result also has page: title, url, capturedAt (ISO 8601), language (the page's own lang, or null) and headings (level and text, in order), plus markdownPath where the .md was written, or markdown with the text itself for inline. For only, path is the .md. An extension older than this feature sends no text, and the call fails with a sentence saying to update it.

With textOptions, page.scope says whether the main content was found and written, page.cut (when maxChars left blocks out) gives the whole text's length and how many blocks went, and each heading in page.headings has offset and chars: where its section starts in the text and how long it is. With an extension older than this, the call fails with a sentence saying to update it.

Given urls, the answer is results, one per page in the list's order: that page's result with ok: true and its url, or ok: false with an error carrying a code and a message. The pages are captured one after another, so a list takes as long as its pages together, and a page that fails does not stop the rest. written and failed count them. The call is an error as a whole only when nothing could be scanned at all: no pairing, no browser, no daemon. For more than 50 addresses, call it again with the rest, or use the command line's --urls.

diff_captures

Says what changed between two captures of a page that scan_page wrote: two Markdown files a passage at a time, or two PNGs pixel by pixel, outlining the changed regions on the newer one. For PDFs, scan with markdown: "beside" and pass the PDFs or their .md files.

Returns kind (text or visual) and changed. For text: added and removed line counts, hunks, the same as a unified diff, each side's source and captured from its front matter, and sameAddress; with section, section says where each side has it (line, heading), or null for a side without it, which counts as a change. For pictures: regions (x, y, width, height in pixels of the newer capture), changedFraction, both sides' sizes, and path, the outlined picture, or null when nothing changed. No browser is needed: both captures are already on disk.

The picture comparison is by position, so content that moved down the page shows as changed below the move; the text comparison sees through that. JPEGs are refused.

extract_design

Reads a page's design, or several pages', in its light and its dark scheme and writes tokens.json, tokens.css, tailwind.preset.js, audit.md, contrast.md and extract.json into a directory, and components.json, components.md and catalog.pdf with components, as page-scanner design does.

Returns the directory and each file's path, the tokens counted by group, how many the page declared, unusedDeclared (declared colors nothing was drawn in, left out), the contrast pairs (pairs, failing below AA, and overImage, text over a gradient or picture left to the eye), nearlyEqual, onePage (values only one page uses) and darkDiffers. pages lists every page with ok and, when it was not read, message; in a crawl, landed and left say where a page went and why it was left out, and crawl says whether robots.txt was read and which links were passed over, by reason. With components, components counts the components and their variants. Inferred tokens are numbered by use, not named, so naming them is left to the agent.

storyboard_page

Scrolls a page from its top at 1.25 screens a second and records how it behaves as a storyboard, as page-scanner storyboard does: a PDF with an overview and a page for each scene with six frames from its entry to its exit, a .json beside it with every scene measured, and the frames as JPEGs in a .frames folder. It is the extension's Scan scrolling, asked for by an agent rebuilding a page's scroll animations.

Returns path, specPath and framesPath, and storyboard, the JSON itself: the address, the viewport, the page's height, the pace, the color scheme, the libraries seen, the page's feel in a sentence, ended (done, stopped, input or detached) and cut, and every scene with its scroll range, kind (pinned, motion or scroll), title, moves (what moved, a sentence each), text, media, feel (on the scenes that moved most) and keyframes, each with its cell (0 to 5), scroll y, file and absolute path. A plain scroll has no keyframes. It takes up to two minutes, so it answers the way scan_page does for a long capture: a client that asks for progress gets it, and one that does not gets status: "running" and calls again with the same arguments.

check_quotes

Before quoting a page in an answer, checks that each quote or value occurs in the capture's Markdown as written, with the line and the section it is under. No browser and no model: the agent writes, Page Scanner checks. The Markdown's own markup is undone first (escapes, emphasis, link syntax, list and quote prefixes, table pipes) and line wrapping is ignored; a quote never runs across two blocks, and a word or number has to stand apart (4.99 is not in $14.99).

Returns checked, found, missing and allFound, and a result per quote: match (exact, loose or none), count, and occurrences, each a line and a section ("Pricing › Pro", or front matter). A quote not found has near: the longest start of it the capture has, where, and how each goes on from there, so the quote can be put right. The same check is page-scanner check on the command line.