Browsing

Browsing with Chrome Direct

Overview

Control Chrome via DevTools Protocol using the use_browser MCP tool. Single unified interface with auto-starting Chrome.

Announce: "I'm using the browsing skill to control Chrome."

When to Use

Use this when:

• Controlling authenticated sessions
• Managing multiple tabs in running browser
• Playwright MCP unavailable or excessive

Use Playwright MCP when:

• Need fresh browser instances
• Generating screenshots/PDFs
• Prefer higher-level abstractions

Auto-Capture

Every DOM action (navigate, click, type, select, eval, keyboard_press, hover, drag_drop, double_click, right_click, file_upload) automatically saves:

• {prefix}.png — viewport screenshot
• {prefix}.md — page content as structured markdown
• {prefix}.html — full rendered DOM
• {prefix}-console.txt — browser console messages

Files are saved to the session directory with sequential prefixes (001-navigate, 002-click, etc.). You must check these before using extract or screenshot actions.

Credential-shaped pages: when a page shows a token or secret (Slack xoxb-/xapp-, GitHub ghp_/github_pat_, 1Password ops_/A3- keys, otpauth:// seeds, or any element marked data-sen-secret), no files are written for that action and the response says ⚠️ Page shows credential-shaped content; auto-capture and DOM output suppressed. with only metadata. Token-shaped values in extract/eval output are replaced by [REDACTED credential-shaped], and screenshot refuses. Capture secrets with a credential broker. On a token-shaped page, use eval only for value-blind queries (e.g. "is the token field present?"). On a page with any data-sen-secret element, eval refuses outright, and extract/attr refuse for the marked element or strip it from HTML/markdown output. The marker scan covers the top document, open shadow roots, and same-origin iframe/frame/object/embed; it does not see closed shadow roots or cross-origin frames. Markup inside an <iframe srcdoc="…"> attribute is not stripped: extract html and attr srcdoc return it verbatim.

When to mark a secret: mark its element with set_attr (data-sen-secret) as soon as the action that revealed it returns, before any other action on that page, then capture the value with the credential broker. Marking only affects later actions. The capture files the revealing action already wrote (e.g. 003-click.html/.md) still hold the value and are not deleted, so don't read them back. This is an accident guard for a cooperating agent, not a security boundary. SUPERPOWERS_CHROME_ALLOW_CREDENTIAL_CAPTURE=1 turns all of this off.

Separately (and unconditionally): a plain password or one-time code has no shape the check above recognizes, so a page's own handler mirroring one into an attribute, a hidden input, or visible text would otherwise land in the .html/.md/-diff.txt files. Whatever gets written is scrubbed on an inert clone (document.implementation.createHTMLDocument + importNode, never cloneNode on the live document): value and every data-*/aria-* attribute are stripped from input[type="password"] (case-insensitive; remembered on that same element after a "show password" toggle flips it to type="text"), any field whose autocomplete contains current-password, new-password, one-time-code, cc-number, or cc-csc (case-insensitive substring), and data-sen-secret-marked elements. A field with none of those signals is still caught if its own live .value exactly matches one of its OTHER attributes (a self-mirror) -- the case for Google's 2-step verification totpPin field, which has no recognized type or autocomplete but copies the typed code into data-initial-value itself. Self-mirror detection skips inputs nobody types into (submit, button, reset, checkbox, radio, hidden, image) and ignores attributes that name or label a field (type, name, id, aria-label, title, placeholder, for, class). The live .value of every field found any of these ways, plus its HTML-entity-escaped forms (including  ), is also replaced with [REDACTED] wherever it occurs in the output -- any attribute, on any element, not just the field's own value -- if it is at least 4 characters long (3 for cc-csc). Dialog captures are redacted with the values from the most recent page capture; a value typed in the same action that opened the dialog is not. Redaction is a plain string replace, so a secret that also appears as ordinary page text (a password of password) replaces that text too; the artifact is degraded but nothing leaks. Self-mirror detection can also flag a non-secret typeable field whose value happens to equal one of its other attributes (a search box synced into data-query), with the same effect. The live page is never touched.

Known gaps (the secret is written to disk in clear): split one-digit OTP boxes feeding an aggregate hidden input (each digit is below the length floor, so the aggregate leaks); a field cleared on submit whose mirror remains; a "show password" toggle that swaps in a new element instead of changing type; a self-mirrored value shorter than the length floor (the field is still found, but its value is too short to substring-redact). Screenshots are pixels, not scrubbed text: a visible (type="text") typed value can appear in a .png.

Default secret-pattern detection (no marking, no click required): everything above needs a recognized type/autocomplete, or a value already mirrored somewhere -- neither helps a secret already in the DOM at page load, before any action (and so before any marking) has run. Real case: Slack's 2FA setup page puts the TOTP seed in a hidden #init_key_code element from load, so the very first navigate's auto-capture would otherwise leak it. Any element whose id/name/class/autocomplete/aria-label contains secret, totp, otp, 2fa, mfa, key_code, seed, recovery, backup_code, api_key, or token as a WORD (split on -/_/camelCase/whitespace, not a raw substring -- "otp" no longer matches inside "footprint", "token" no longer matches inside "token-list") is redacted on sight: a value-bearing match (input/textarea/select) joins the same value-collection/redaction pass as above; a leaf match (no element children -- a <span>/<div>/<code> holding the secret as text) has its own text blanked directly by element reference, regardless of length, never via a page-wide substring replace (that used to corrupt unrelated text sharing the same word, e.g. Prism's class="token keyword" spans turning every class= attribute into [REDACTED]=); a wrapping container (has element children) is fully blanked -- every descendant's text is blanked too, not just its own attributes, so a secret split across child elements (<div id="totp-secret">Key: <code>SEED</code></div>, a <ul class="recovery-codes"> list) no longer leaks -- but only when the match is STRONG (an exact compound like totp-secret/recovery-codes/api-key/private-key, not a bare single word) and the container's text is 2000 characters or fewer; a WEAK match (bare token/secret, e.g. a design-system class sn-token-provider or a docs <section id="module-secrets">) does not wholesale-blank -- that wiped real page content (a "Get started" paragraph) in testing -- though any descendant that independently matches on its own is still found and blanked regardless. A button/link (and anything nested inside one) is never text-blanked even when matched (its own text is a UI label, e.g. copy-seed-btn's "Copy", not the secret) -- but its own attributes ARE still stripped (fixed: an earlier version returned before stripping them, so a copy button's data-clipboard-text leaked). label/summary/legend/option are no longer exempt either -- their own text is blanked and recursed into, since a <label> wrapping a matched leaf is routinely the only place the secret renders at all. An otpauth:// URI is redacted unconditionally wherever it appears (href/src/data-*/text/alt/title), independent of the word-boundary match. The .md artifact gets the same redaction independently (it walks the live DOM directly, not the HTML clone), including a matched element sitting INLINE inside an emitted block (<p>Key: <code id="totp-secret">SEED</code></p>) by walking each block's own subtree instead of flattening it via textContent. A bare token is broad enough to also catch an ordinary CSRF token field -- accepted, since this only touches the disk copy, never the live page.

URL-based suppression: independent of markup, a page whose URL matches /2fa, /two_factor (unanchored -- GitHub/GitLab/1Password's real 2FA routes all need the open suffix, same as Slack's), /two-step, /mfa (segment-anchored, unlike the others -- an unanchored /mfa also matched a docs page merely about mfa), /totp, /security/keys, security-keys, /recovery-codes, /backup-codes, login-verification, or security-info has its html/md capture AND its screenshot -- including an explicit screenshot action, not just auto-capture -- suppressed outright, same as a credential-shaped page but with its own distinct message (an earlier version always said "page shows credential-shaped content" for an explicit screenshot refusal, even when the real reason was the URL). /2fa//totp/security-keys/security-info require a word-boundary on the open side -- /2fast, /totpal and /security-keyset-docs no longer match just because they start with the same letters. Matched against the URL's pathname and hash (never its query string, which over-suppressed an ordinary /login?next=/settings/2fa redirect) -- the hash is included so a hash-routed SPA route (e.g. AWS's #/security_credentials/mfa) still matches -- and only for a real page location (http(s):, file:); never for data:/blob:/about: URLs, whose "pathname" is not a path (a data: URL's pathname IS its percent-encoded body, so a page merely mentioning a sensitive word in its own title would otherwise false-match on itself). Catches a secret with no recognized shape or attribute naming at all (a bare seed, a QR code) purely by the page it's on. SUPERPOWERS_CHROME_ALLOW_CREDENTIAL_CAPTURE=1 disables this (for auto-capture and an explicit screenshot alike). SUPERPOWERS_CHROME_SENSITIVE_URL_PATTERNS adds extra comma-separated regexes to the defaults (never replaces them).

The use_browser Tool

Single MCP tool with action-based interface. Chrome auto-starts on first use.

Parameters:

• action (required): Operation to perform
• selector (optional): CSS or XPath selector for element operations
• payload (optional): Action-specific data (string or object)
• timeout (optional): Timeout in ms for await operations (default: 5000)

Active tab: Every action operates on the current activeTab. Use switch_tab to change it.

Actions Reference

Navigation

• navigate: Navigate to URL

  • payload: URL string
  • Example: {action: "navigate", payload: "https://example.com"}

• await_element: Wait for element to appear

  • selector: CSS selector
  • timeout: Max wait time in ms
  • Example: {action: "await_element", selector: ".loaded", timeout: 10000}

• await_text: Wait for text to appear

  • payload: Text to wait for
  • Example: {action: "await_text", payload: "Welcome"}

Interaction

• click: Click element

  • selector: CSS selector
  • Example: {action: "click", selector: "button.submit"}

• type: Text input

  • selector: Optional — clicks to focus first
  • payload: Text to type (\t=Tab, \n=Enter)
  • Example: {action: "type", selector: "#email", payload: "user@example.com"}

• double_click: Double-click element (fires dblclick event)

  • selector: CSS selector
  • Example: {action: "double_click", selector: ".item"}

• right_click: Right-click element (fires contextmenu event)

  • selector: CSS selector
  • Example: {action: "right_click", selector: ".row"}

• select: Select dropdown option

  • selector: CSS selector
  • payload: Option value(s)
  • Example: {action: "select", selector: "select[name=state]", payload: "CA"}

• keyboard_press: Press special keys (Tab, Enter, Escape, Arrow keys, F1-F12)

  • payload: Key name (string) or {"key": "Tab", "modifiers": {"shift": true, "ctrl": false, "alt": false, "meta": false}}
  • Example: {action: "keyboard_press", payload: "Tab"}
  • Example with modifiers: {action: "keyboard_press", payload: {"key": "Tab", "modifiers": {"shift": true}}}

Mouse Actions (CDP-Level)

These use CDP Input.dispatchMouseEvent, bypassing synthetic event restrictions.

• hover: Move mouse over element (CSS :hover, tooltips, menus)

  • selector: CSS selector
  • Example: {action: "hover", selector: ".menu-trigger"}

• drag_drop: Drag element to target (native drag-and-drop via CDP)

  • selector: Source element
  • payload: Target selector or JSON coordinates {"x":N,"y":N}
  • Example: {action: "drag_drop", selector: ".card", payload: ".column-2"}

• mouse_move: Move mouse to coordinates

  • payload: JSON {"x":N,"y":N} (optional: steps, fromX, fromY for smooth movement)
  • Example: {action: "mouse_move", payload: "{\"x\":100,\"y\":200}"}

• scroll: Scroll via mouse wheel events

  • payload: Direction (up/down/left/right) or JSON {"deltaX":N,"deltaY":N}
  • selector: Optional — scroll within element
  • Example: {action: "scroll", payload: "down"}

File Upload

• file_upload: Set files on input[type=file] elements (can't be done via JavaScript)
  • selector: File input element
  • payload: File path or JSON {"files":["/path/a.pdf","/path/b.jpg"]}
  • Example: {action: "file_upload", selector: "#upload", payload: "/tmp/doc.pdf"}

Extraction

• extract: Get page content

  • payload: Format ('markdown'|'text'|'html')
  • selector: Optional - limit to element
  • Example: {action: "extract", payload: "markdown"}
  • Example: {action: "extract", payload: "text", selector: "h1"}

• attr: Get element attribute

  • selector: CSS selector
  • payload: Attribute name
  • Example: {action: "attr", selector: "a.download", payload: "href"}

• set_attr: Write-only attribute setter, restricted to EXACTLY two attribute names — data-sen-nonce and data-sen-secret itself (nothing else, not any other data-*/aria-* name; see skills/browsing/lib/set-attribute.js's ALLOWED_ATTRIBUTE_NAMES constant)

  • selector: CSS or XPath selector
  • payload: {"name": "data-sen-nonce"|"data-sen-secret", "value": "..."} (no bare-string form — needs both fields)
  • data-sen-nonce resolves to the single first-VISIBLE match, the same way extract/click/type resolve a selector, and refuses if that element already carries data-sen-secret. data-sen-secret instead marks EVERY element the selector matches, hidden duplicates included, and can never remove or weaken an existing mark (re-marking an already-marked element is a no-op, not a refusal) — this is the write path for the marker itself.
  • Unlike every read action, set_attr is NOT blocked by a data-sen-secret element existing elsewhere on the page — it takes no caller JavaScript, and its only output is ok/no element matched/refused, which acts as a limited prefix oracle (see below). Use it instead of eval to write onto a page that already has a captured secret (e.g. stamping a credential-broker nonce onto an unmarked digit-input box next to a just-captured TOTP seed, or marking the seed's element in the first place).
  • Why so narrow: page JS and frameworks routinely read arbitrary data-*/aria-* attributes and wire them to behavior (data-action, data-href, aria-controls, and more a hostile page could invent), so a prefix allowlist is not guaranteed inert. Widening past these two names is a deliberate, separate change. Its ok/no element matched/refused: target element is marked responses differ by outcome, which is itself a limited prefix oracle over page content for a caller who varies the selector and watches which result comes back — a known, accepted limitation, not something this guards against.
  • Example: {action: "set_attr", selector: "#code-input-0", payload: {"name": "data-sen-nonce", "value": "opaque-nonce"}}

• eval: Execute JavaScript

  • payload: JavaScript code
  • Example: {action: "eval", payload: "document.title"}
  • Refuses outright (no value-blind exception) while any element on the page carries data-sen-secret, checked live at the moment of the call — see set_attr above for the write-only escape hatch. This is an accident guard, not a security boundary: eval runs in the same JS realm as the marked element, so it can already read the value directly, exfiltrate it via fetch()/window.name/storage, or erase the marker with removeAttribute as its own last step — none of which this check can catch, by design. Don't mark an element and then eval on that page expecting the value to stay contained; after marking, use the credential broker for the value and set_attr for writes.