NexusBrowser

Your real Chrome, driven by AI agents in background tabs. NexusBrowser lets Antigravity, Claude and Codex use the user's own Chrome (their sessions, their extensions) through one shared local bridge and one MCP server, nexus-browser. Each agent works in its own background tabs, grouped under its name, with real (trusted) mouse and keyboard events sent through the Chrome DevTools Protocol, while the user keeps working in their own tabs.

Windows 10/11 · Chrome 116+ · Python 3.14 · MIT licence

Part of the Nexus family: NexusWindows (Claude · Codex · Antigravity), NexusBrowser (repo), NexusAndroid (repo). Project page: julienpiron.fr/nexusbrowser.

Why

Browser automation usually means a separate headless browser (no sessions, no extensions) or screen-level computer use that moves the user's cursor and brings the window to the front. NexusBrowser works inside the browser the user already uses, without activating the agent's tab or window, and keeps the guard rails in the extension and the bridge rather than in the agent: tab ownership, an allowlist of DevTools commands, masked secrets, an emergency stop.

• Antigravity: its own browser agent runs in a separate Chrome profile, without the cookies or sign-ins of the user's browsing (Antigravity docs). NexusBrowser gives it the user's real Chrome.
• Claude and Codex already have their own Chrome extension (Claude in Chrome, the Codex Chrome extension). NexusBrowser complements it where a bridge with the DevTools protocol helps: JavaScript dialogs (browser_handle_dialog), detailed network capture with response bodies, elements inside cross-origin iframes, content and Google Docs/Sheets/Slides exports, and downloads.

Until 2.4.0 it was published here as "Antigravity Browser Bridge", for Antigravity only. Since 2.6.0 it is a neutral bridge shared by several agents.

Pieces

Tools

52 MCP tools.

Install

Requirements: Windows 10 or 11, Chrome, uv, Python 3.14.

git clone https://github.com/pironjulien/NexusBrowser.git
cd NexusBrowser
pwsh -NoProfile -File install.ps1

• Python environment, token, logs, screenshots and exports live in ~\.nexus\nexus-browser (outside the clone).
• Scheduled task Nexus Browser Bridge with no trigger: the bridge starts on demand (first agent call, schtasks /Run).
• Native host com.nexus.browser_auth declared to Chrome (HKCU): it computes the extension's proofs, the token never reaches Chrome.
• MCP server nexus-browser registered for Claude Code, Codex and Antigravity when they are installed, each with NEXUS_BROWSER_AGENT set to its name (default tab group).
• When the bridge has to relaunch Chrome, it uses the Default profile; extra Chrome flags (language...) can be added with the NEXUS_CHROME_EXTRA_ARGS environment variable (allowlist: --lang, --accept-lang, --variations-override-country, --disable-features=Glic…, --force-dark-mode; anything else is ignored).
• Migration from the former "Antigravity Browser Bridge": its task (same ports) and native host com.nexusagy.bridge_auth are removed.

Then load extension\ once in chrome://extensions (developer mode, "Load unpacked") and remove the old "Antigravity Browser Bridge" if present. The extension ID is fixed by the public key in manifest.json (blcgddmjdkobmgakjffjpgiicjlflgjf), the same on every PC and whatever the folder; it is the only origin the bridge and the native host accept.

The MCP server from PyPI

Once the bridge and the extension are installed (above), any MCP client can also start the server straight from PyPI, without the clone:

claude mcp add nexus-browser --scope user -e NEXUS_BROWSER_AGENT=Claude -- uvx nexus-browser

Without the bridge, the server still starts and lists its 52 tools, and every call answers with a clear error that points to the one-time install.

Security (held by the extension and the bridge, not by the agent)

• Mutually authenticated bridge (nexus-browser/1): HMAC proofs over both nonces, then every message signed with a per-connection key and a strict sequence number. HTTP requires X-Nexus-Browser-Token and refuses any browser origin.
• Tab ownership: every bridge client identifies itself (agent + session headers, forwarded by the bridge in the signed channel). A tab opened by browser_new_tab belongs to that session (chrome.storage.session); any page command on another tab (the user's, another agent's) is refused. A tab of the user only through browser_adopt_tab(tab_id, user_asked=True). tab_id is mandatory on every page tool. A session whose MCP process is gone (or silent for 1.618 x 60 s; servers beat every 37 s) leaves orphan tabs: the same agent can reclaim (browser_adopt_tab, also automatically at startup), close or release them; never another agent.
• CDP allowlist for browser_cdp_send; every extension script runs in an isolated world (only browser_eval, the YouTube transcript and browser_cdp_send(..., world="page") touch the page's world).
• Secrets: secret headers, URL parameters and JSON/form body fields masked (pagination tokens and LLM token counters kept); cookie values only for local development origins (include_values=True); uploads refused by the bridge itself from secret folders (normalized paths). The value of a password, one-time code or payment card field (type=password, autocomplete current-password, new-password, one-time-code, cc-number, cc-csc, cc-exp) never leaves the page: browser_content (html mode included), browser_vision_inspect, browser_find, browser_axtree and the value read back by browser_fill return [masked value] instead.
• browser_eval is full access to the page, nothing masked: like browser_cdp_send(Runtime.evaluate) (both worlds share the document), it reads what the page reads (non-HttpOnly cookies through document.cookie, localStorage, sessionStorage, IndexedDB, form fields) and its fetch carries the user's session. The masking above covers the dedicated tools, not scripts nor the raw Chrome replies of browser_cdp_send (DOM.getOuterHTML, Accessibility.getFullAXTree...): kept on purpose, said plainly in the tool, the skill and the server instructions. Page.reload with scriptToEvaluateOnLoad is refused (script injected at load).
• Outward gestures need the user's accord for the exact action (2.12.0): before a click (browser_click, browser_click_mark, browser_semantic_click), Enter / Control+Enter / Space (browser_press_key, keys of browser_fill and browser_semantic_fill), a typed line break (browser_type_keys) or a drag, the server reads the trigger without any gesture, through the bridge's existing read paths (find for locators, and a read-only script, client/trigger_probe.js, evaluated by cdp-send in the extension's isolated world): no new extension action, so it works with the extension already loaded. It reads the clicked control, or the submit buttons of the focused field's form. When the shared policy (client/garde-fous.json) sees an outward word (send, pay, buy, order, publish, post, call, transfer, subscribe, book...), the gesture is refused with a summary of the exact action (target, recipients, form content without sensitive values, attachments, page origin and path) and a one-time token valid 10 minutes. The agent shows the summary to the user and, only after their yes, sends the same call again with confirm=true and accord="<token>". Any change between the accord and the gesture voids it; confirm=true alone is never enough; a token read in a page is unknown. Neutral links and buttons stay free. The check fails closed only when the read really fails (script error, focus in a cross-origin frame, form of a field aimed at by Enter unreadable). Page text outputs (browser_content, browser_axtree, browser_export, Google and YouTube exports, network bodies) carry untrusted_note when the text looks like instructions aimed at an AI (a signal only). The CLI runs the same check but cannot carry a token from one command to the next (one process each): it refuses outward gestures with the summary and points to the MCP server; every other gesture stays free.
• WebMCP tools (sites that expose their own tools to agents through document.modelContext, native in Chrome 149+ for sites in the origin trial): browser_webmcp_tools lists them and browser_webmcp_call runs one through Chrome's own getTools() / executeTool(), evaluated by cdp-send in the extension's isolated world (the page cannot replace them; no new extension action, no Chrome flag). The call goes through the same accord: a tool the page marks consequentialHint, or whose name, title, description or form submit button carries an outward word, is refused with the exact call (tool, origin, input with sensitive values masked but bound to the token, page) and a one-time token. A tool marked readOnlyHint runs freely. The tool runs only if its definition did not change since the check. Descriptions and outputs are page text (untrusted_note).
• Paths that do not go through the accord: browser_eval and browser_cdp_send (Runtime.evaluate, Input.dispatch*) keep full page access by design and can click or submit without this check; the tool descriptions and the server instructions tell the model never to use them for an outward action. Bare coordinates of browser_drag_and_drop are not read either.
• Emergency stop (Ctrl+Alt+Shift+Esc, ~\.nexus\stop): commands already running are cancelled (typing, drag, waits), debuggers detached, every action refused for 120 s. The same file protocol is shared with the NexusWindows agents.
• Who drives what: each agent's tabs sit in a Chrome tab group named after it; browser_release_tab gives a tab back. The toolbar icon opens a read-only popup: bridge state and tabs per agent.
• Whoever holds the bridge token controls the browser session: keep ~\.nexus\nexus-browser\bridge-token private.

Checks

pwsh -NoProfile -File test.ps1         # offline: ruff, Pyright, pytest (real bridge on ephemeral ports), Node tests
pwsh -NoProfile -File test.ps1 -Full   # plus real validation in the user's Chrome: --mouse then --full-live
pwsh -NoProfile -File verify.ps1 -Browser  # task, native host, bridge reached, loaded extension up to date

test.ps1 uses the environment created by install.ps1.

Updating an existing installation also checks the extension's loaded version/features and reloads it when needed, then reloads the supervised Browser MCP worker. The installation fingerprint covers client and extension sources. An update defers while an agent owns browser tabs; user tabs are never driven or activated by the refresh. If Chrome is closed, it loads the updated extension files on its next launch; that state is not a live validation.

Privacy

NexusBrowser runs entirely on your PC: a Chrome extension, a local bridge bound to 127.0.0.1 (authenticated by a local token) and an MCP server. It has no account, no telemetry, no analytics and no remote server; nothing is sent to the author or to any third party.

• What it reads: in the tabs the agent opened (or a tab of yours you explicitly let it adopt): page text and accessibility tree, screenshots, console logs, network requests (headers and secret fields masked), cookies when browser_get_cookies is called. Password, code and card values are masked in reads and drawn as dots in screenshots and PDFs (every frame, cross-origin ones included); console logs are masked like network bodies; browser_eval and Runtime.evaluate through browser_cdp_send do have full access to the page.
• Network: pages are loaded by your own Chrome with your own sessions, as if you browsed them. Exports (Google Docs, Sheets, Slides, YouTube transcripts) are fetched from those sites in your session.
• Where it goes: tool results are returned to the AI client you connected (Antigravity, Claude, Codex…), which may send them to its model provider under its own privacy policy.
• What it stores, under ~/.nexus/nexus-browser/: the Python environment, the bridge token, bridge.log (connection and emergency-stop events, rotated at 1 MB, three backups), install.log, screenshots (at most 50, each kept under 24 hours) and the exports you ask for. The extension keeps tab ownership in Chrome's session storage, cleared when Chrome closes.
• Removing it: remove the extension from Chrome, delete the scheduled task Nexus Browser Bridge and the registry key HKCU\Software\Google\Chrome\NativeMessagingHosts\com.nexus.browser_auth, then delete ~/.nexus/nexus-browser/.

Questions or concerns: open an issue on this repository.

Limits

• Windows and Chrome only (Manifest V3, chrome.debugger). While the debugger is attached, Chrome shows its "is being debugged" banner on the agent tab. Protected pages (chrome://, the Web Store) cannot be scripted.
• Tab ownership is per MCP process: sub-agents of one conversation share it.
• The YouTube transcript reads the page's own player (page world).
• The outward-action check reads the trigger just before the gesture: a page that rewrites its form or its button in between is not seen. It recognizes <form>, dialogs and the nearest block holding a field and a button (at most 3 fields, like a chat composer); recipients shown as custom chips are read only when they carry an email attribute. The token proves that this exact action was described before running unchanged, not who said yes: the agent stays the only channel between the user and the tool.
• WebMCP: the page's own readOnlyHint and consequentialHint are trusted as declared, and tools of a cross-origin iframe are listed only when that frame exposes them to the top page (exposedTo), as the spec requires. Chrome's CDP WebMCP domain (used by Chrome DevTools MCP) is not used: it needs Chrome started with --enable-features=DevToolsWebMCPSupport.
• Code comments are in French.

Licence

MIT, see LICENSE. Copyright (c) 2026 Julien Piron.