NexusBrowser

Types text with real keyboard events (trusted typing) into an input, textarea or contenteditable (rich editors such as Gmail, Notion), in the main frame or an iframe (cross-origin included). Flow: focus by a real CDP click on the field (or DOM focus); replace=True: real Ctrl+A then text (Input.insertText: trusted beforeinput/input); replace=False: append at the end; per_key=True: key by key (trusted keydown/keyup, slower); keys: keys or shortcuts sent afterwards (["Enter"], ["Control+Enter"], ["Tab"]). The value is read back: verified (true/false), value ("[masked value]" for a password, a on…

Fills several fields in ONE call, in order, with the same gestures as browser_fill, browser_select_option and browser_check (trusted typing), then READS BACK the whole form without any gesture. Each field: "mark" (number from browser_interactive_map, call it first) and one value: "value" (text), "option" (visible text of a <select> option) or "checked" (true/false). E.g. [{"mark": 1, "value": "Marie"}, {"mark": 4, "option": "Lille"}, {"mark": 6, "checked": false}]. Everything is validated before the first gesture. Stops at the first unconfirmed field (error, dialog open, value refused or reset…

Lists, without acting, the elements matching a locator (see browser_click) in all frames. Response: count, unique (an action would be unambiguous), engine ("ax" = Chrome accessibility tree, "dom"), frames, and for each candidate: role, name, tag, id, frame (frame, crossOriginFrame), visible, enabled, checked/expanded/selected/level, value, box (main viewport coordinates).

Lists the cookies of the tab (or of urls) via CDP Network.getCookies: name, domain, path, expiry, httpOnly, secure, sameSite. Never their values by default (session secrets). include_values=True: values included ONLY for a local development origin (localhost, 127.0.0.1, [::1], *.localhost, *.test: the tab, or each of the urls) and only for cookies of those domains; clear refusal for any other origin. An action, not a read: refused during the emergency stop and never retried automatically.

Puts tabs of this session into a named (default: the agent's name), colored tab group. Supported colors: 'grey', 'blue', 'red', 'yellow', 'green', 'pink', 'purple', 'cyan', 'orange'. Response: updated=false (and warning) if Chrome did not apply the title, color or collapsed state.

Answers the JavaScript dialog (alert, confirm, prompt, beforeunload) open in tab tab_id (never another tab's dialog). No dialog is ever accepted automatically: when a command opens one, its response reports it (dialog / pendingDialogs fields: type and message, for the targeted tab only). Read the message, then choose: - accept=True: OK / Leave page; accept=False: Cancel / Stay. - prompt_text: text entered in a prompt dialog before accepting.

Hovers an element with a real CDP mouse move (isTrusted events), tab left in the background. Reveals dropdown menus, submenus, tooltips and hidden popovers. Response: method ("cdp" or "js_fallback") and trusted_hover. An element covered by an overlay is reported as an error; nothing else is hovered. Targeting, one of: - mark: badge number from browser_interactive_map or vision_inspect (e.g. mark=5) - selector: CSS selector or mark hash (e.g. "#5", ".dropdown-trigger") - text: visible text of the element (e.g. "Options") - locator: role and label locator (see browser_click), takes precedence.

Extracts a condensed map of the visible interactive elements ([#1] button, [#2] input...). Recommended discovery tool: 85% to 95% fewer tokens than raw HTML or AXTree. Gives for each visible element: its mark number, tag, ARIA role, text label, optimized CSS selector and box_2d coordinates [ymin, xmin, ymax, xmax] (0-1000). Lets you chain clicks and fills directly with browser_click(mark=N) or browser_fill(mark=N, value=...).

Lists the tabs open in Chrome (id, title, URL) and, for each, the agent driving it (owner), whether it belongs to THIS session (mine) and its state: "mine" (this session), "orphan" (orphaned by my agent: dead session, take it back with browser_adopt_tab, or close / release it), "other" (another agent or another live session) or "user" (the user's tab, no agent). Only mine=true tabs accept page commands.

Returns the stream of errors and messages from the tab's DevTools console.

Navigates tab tab_id to the URL (http, https, data: or about:blank; never file:// or chrome://). url="back" or "forward": previous or next page of this tab's history (the browser's back and forward buttons).

Starts (or extends) the tab's network recording for duration_s seconds (at most 600): the debugger stays attached until then, even without commands. Stop earlier with browser_network_stop. Without a requested recording, the debugger is detached after 30 s of inactivity (log kept, bodies lost).

Network requests recorded for the tab (ring buffer of 500), most recent last: url, method, status, type (Document, Fetch, XHR, Script...), mimeType, size (bytes received), duration_ms, state, failed and errorText, request_id (for browser_network_response). Filters: filter (URL fragment or "/regex/"), method, status (200, "4xx", "400-499"), resource_type, failed_only. include_headers=True adds headers; Cookie, Set-Cookie, Authorization and any header whose name suggests a token, key, session, signature or password are always masked, as are sensitive URL parameters (access_token, code, key, sig…

Response body of a request from browser_network_requests (Network.getResponseBody), truncated to max_bytes (truncated). Secret fields masked by the extension, at any depth: JSON (access_token, refresh_token, id_token, token, password, secret, api_key, client_secret, authorization, session...) and forms (a=1&token=...); masked = number of masked values. A binary body is reported (binary, size) and returned as base64 only with include_binary=True.

Stops network recording and detaches the debugger from the tab (the log already captured stays readable).

Opens a new Chrome tab at the given URL (or about:blank) and returns its id: the tab_id to pass to every following command. The tab belongs to THIS agent session (enforced by the extension). By default in the background (background=True) without stealing the user's focus, in the tab group named after the agent (Antigravity, Claude or Codex); group="": no group. Only http, https, data: or about:blank URLs (never file://, chrome://).

Generates a high-fidelity vector A4 PDF export of the tab's page.

Presses a real key or shortcut via CDP on the focused element: Enter, Escape, Tab, Backspace, ArrowDown, F5, or a combination "Control+A", "Shift+Tab", "Control+Shift+K" (key, code, windowsVirtualKeyCode, modifiers). Enter, Control+Enter or Space toward an outward control (a focused send/pay/order button, or the submit button of the focused field's form): accord required, as for browser_click.

Gives the tab back to the user without closing it (ownership removed, debugger detached). For an adopted tab, or an orphan tab of my agent.

Reloads the page of tab tab_id.

Screenshot of the tab, of the whole page (full_page=True), or of one element cropped to the pixel (PNG), even if the page has scrolled. By default, the image is written to ~/.nexus/nexus-browser/screenshots/ and only its path is returned (output_path): no bulky base64 in the response. Do not open a capture larger than 500 KB with an image reader. Targeting: - selector: CSS selector of the element to crop (e.g. "#dashboard-card") - mark: mark number to crop (e.g. mark=4) - output_path: absolute PNG path (optional, otherwise the local cache) - full_page=True: the whole scrollable page, beyond t…

Scrolls the web page by y pixels down (down) or up (up).

Selects an option of a <select> (targeted by locator, e.g. {"label": "Plan"}, or selector) by value or by visible text, then fires input and change events (like Playwright selectOption). A custom list (role=combobox without <select>) is driven with browser_click on the list, then on {"role": "option", ...}.

Activates and brings to the front this session's tab (steals the user's focus: only at the user's request).

Types text key by key via CDP Input.dispatchKeyEvent (trusted keydown/keypress/input/keyup), for fields that require real keys (input masks, page shortcuts). No target: the focused element; with locator or selector: focus first (real click), append at the end, value read back (verified). keys: keys or shortcuts sent afterwards (["Enter"]). The emergency stop interrupts typing in progress. A line break in text, or a submitting key, toward an outward button of the field's form: accord required, as for browser_click (the typed text is part of it).

Uploads one or more files (e.g. .aab, .apk, .zip) without opening or blocking on the native Windows file dialog. Refused (by the bridge itself) for any file in a secrets folder: ~/.nexus, ~/.nexusagy, ~/.ssh, ~/.codex, ~/.claude, ~/.gemini, ~/.android, ~/.git-credentials, ~/.config/gcloud, %APPDATA%/gcloud, Chrome profile, the user's secrets vault... (normalized paths: \\?\, \\localhost\C$, links and junctions resolved).

Places numbered Set-of-Mark (SoM) badges on interactive elements and captures the view for visual analysis; the numbers are then used with browser_click(mark=N). Returns the list of elements with box_2d [ymin, xmin, ymax, xmax] on a [0-1000] scale. The capture is written by default to ~/.nexus/nexus-browser/screenshots/ (full-resolution PNG, path returned in output_path). inline=True attaches it to the response as an image instead: WebP, long side at most 1568 px and 1.15 megapixels, at most 500 KB; data.image gives its size and css_per_image_px to read a position on it, and each marker its i…

Waits for a download started in the tab (link click...) to finish and returns its path, size (bytes) and type (mime). Candidates: downloads started since the agent's last action on the tab (or since_ms, a timestamp in milliseconds), coming from the tab's origin. A file Chrome flags as dangerous is never accepted by the agent: explicit error. timeout_ms: at most 115000.

Waits until a visible element shows the text (case-insensitive substring) or matches the locator (see browser_find), in every frame; with gone=True, until none is left (a spinner, "Loading...", a closed dialog). Use it after an action whose result appears later: no fixed pause needed. timeout_ms: at most 115000. Response: count of matching elements and elapsed_ms; an error if the timeout is reached.

Waits until the document is settled and complete (document.readyState === 'complete'). timeout_ms: at most 115000 (bridge limit); the response is awaited for timeout_ms + 5 s.

Waits until network traffic settles (0 active Fetch/XHR request for idle_ms). Call it after a click, a submission or a navigation in a single-page application (React, Angular, Vue, Google Cloud / Play Store consoles) before inspecting the DOM. timeout_ms: at most 115000 (bridge limit).

Calls a WebMCP tool exposed by the page (see browser_webmcp_tools) with input, a JSON object matching its inputSchema; origin picks one when several frames expose the same name. Runs the page's own handler through Chrome's executeTool from the extension's isolated world; timeout_ms bounds it (then cancelled). Same guardrail as clicks: a tool the page marks consequential, or whose name, title, description or form submit button says send, pay, order, book, publish, call, transfer... is refused with a summary of the EXACT call (tool, origin, input, page) and a one-time token; resend the same call…

Lists the WebMCP tools the PAGE ITSELF exposes to agents (document.modelContext, native in Chrome 149+ on sites enrolled in the WebMCP origin trial): name, title, description, inputSchema, annotations (readOnlyHint, consequentialHint, untrustedContentHint), origin, frame, and for a declarative tool (an HTML form) its submit buttons and form target. Read without any gesture, from the extension's isolated world. When a page offers a tool for the task, calling it (browser_webmcp_call) is more reliable than clicking through its interface. supported=false: the page exposes none. Names, descriptions…

Transcript of the YouTube video open in the tab (the player's caption tracks, read in the page world: the only export that needs it). lang: wanted language code ("fr", "en"...): manual track first, then auto-generated (track.auto: true), else YouTube's automatic translation (track.translated); omitted = the video's default track. format: "json3" (default) or "vtt". Response: timed_text ("[mm:ss] text" per line), text (plain text), track, available_tracks, title, channel, truncated (max_chars). If YouTube requires its player's token, the player's captions are turned on briefly then restored (e…