Navigate to a URL

Go back to the previous page

Go forward to the next page

Capture accessibility snapshot of the current page. Use this for getting references to elements to interact with. The snapshot also lists all bound tabs so the LLM can pick a different tab via tabId on subsequent calls.

Perform click on a web page

Hover over element on page

Type text into editable element

Select an option in a dropdown

Press a key on the keyboard

Wait for a specified time in seconds

Get the console logs from the browser

Take a screenshot of the current page

List all browser tabs bound to this agent. Returns tabId, label, URL, title, and which one is the agent's active tab. Use this to discover what's available before issuing tool calls that need a specific tab.

Open a new browser tab and bind it to this agent. Optionally provide a URL (will navigate after open) and a human-readable label (the LLM uses the label to refer to the tab in subsequent calls). The new tab is set as the agent's active tab.

Close a browser tab previously bound to this agent. The tab's binding is automatically removed.

Set a human-readable label on a bound tab. The label is what the LLM uses to refer to the tab in conversation and in browser_list_tabs output.

Set which bound tab is the agent's active tab. Tool calls that don't specify a tabId route to the active tab. Use this to switch the agent's focus between bound tabs.

Click a 'Click to copy' button on the page and return the value the page wrote to the clipboard. Use this for Stripe's publishable key / secret key copy buttons, GitHub PAT copy buttons, AWS access key copy buttons, and any other 'click to copy' UI patterns. The content script patches navigator.clipboard.writeText to capture the value as it is written, then returns it to the LLM.

The returned value is plain text (the literal content of the copy button). For secret material, the LLM should pipe it to the user's PQC secrets store (e.g. via the pqc-secrets CLI) rather than persisting it in chat history or unencrypted state.

Paste text into a focused element on the page. If ref is provided, that element is focused first; otherwise the currently focused element receives the paste. Counterpart to browser_copy_to_clipboard — copy a value with the copy tool, then paste it elsewhere. Dispatches a synthetic paste event (clipboardData.setData + 'insertFromPaste') so React/Vue/Angular controlled inputs accept the value. Use for Stripe webhook secret fields, generated API key fields, or any text you need to push into a form.

Wait for a specific text to appear anywhere in the page DOM (case-insensitive substring match). Polls every 500ms until the text is found or the timeout elapses. Replaces blind browser_wait(time) calls. Returns a fresh ARIA snapshot once the text appears. Use after clicking a submit button, or to wait for a Stripe dashboard to finish loading. On timeout, returns an error with the current page text length so you can diagnose what is or isn't rendering.

Read a single HTML attribute from an element referenced by snapshot ref. Use attr: 'href' to get a link's URL (e.g. extract a Stripe payment link from a share button), attr: 'value' for current input values, attr: 'aria-label' for accessibility text, or any custom data-* attribute. Returns the attribute value as a string, or an empty string if the attribute is not set. Cheaper and more precise than a full snapshot when you need one specific value.

Extract the textContent of a single element by ref. Cheaper than a full snapshot when you just need the visible text of one element. Returns the text as a plain string with whitespace normalized. Use to grab a heading, a status message, or a single field's value without paying the snapshot cost.

Run an arbitrary JavaScript expression in the page's main frame context. Returns the JSON-serializable result. Escape hatch for when ARIA snapshots cannot introspect something — React/Vue component state, computed styles, custom data attributes, third-party widget internals, hidden DOM the snapshot can't reach. The result must be JSON-serializable and is capped at 10KB. The expression runs in the page's global scope via indirect eval (sees window, document, etc.) and any side effects DO persist — only call this with code you have reviewed.