Auto Browser

Tracking since Apr 24, 2026.

Changelog

• May 17, 2026
  permissions

  storage, unlimitedStorage, activeTab, tabs, sidePanel, scripting, debugger, offscreen

  storage, unlimitedStorage, tabs, sidePanel, scripting, debugger, offscreen

• May 12, 2026
  description

  An AI agent that actually uses your browser.
  
  Auto Browser is a Chrome side-panel extension that lets an AI agent read and act on any web page under your direction. Describe a task in plain English — filling a form, extracting data, comparing products across sites, running a checkout flow — and the agent sees the page, plans the steps, and does the work. You review and approve each action before it fires.
  
  No scripts to write. No per-site setup. No account required.
  
  
  — — —
  
  WORKS EVERYWHERE
  
  Any website, any time. No per-site integration, no API stubs, no "connect your account" dance. The agent reads the page through standards the browser already speaks.
  
  
  BRING YOUR OWN AI
  
  Four providers, one dropdown. Switch anytime without losing your conversation:
  
  • Chrome Built-in AI (Gemini Nano) — free, on-device, offline-capable. No API key, and nothing leaves your device.
  • Google Gemini — frontier-quality reasoning, direct API using your own key.
  • OpenRouter — access Claude, GPT, Llama, Mistral, and every other model in the OpenRouter catalog with a single key.
  • Local LLM — point Auto Browser at any OpenAI-compatible endpoint you run yourself (Ollama, LM Studio, llama.cpp, and similar).
  
  Your conversation history carries over when you switch.
  
  
  PRIVATE BY DEFAULT
  
  • No analytics. No telemetry. No servers of ours between you and the AI provider you choose.
  • Your API keys stay on your device. We never receive them.
  • Passwords, tokens, and other secret fields are redacted before they can reach the model.
  • Choose on-device Gemini Nano or a local LLM and zero data leaves your machine.
  
  
  SAFE BY DESIGN
  
  • The agent asks before every write — clicks, typing, navigation. Read-only looks at the page don't interrupt you.
  • Banking, government, and healthcare sites are blocked by default. The agent is disabled entirely on those origins.
  • Executing arbitrary JavaScript always prompts, regardless of how broadly you've approved a site.
  • If a page redirects mid-action (login loops, phishing hops), prior approvals no longer apply — the agent has to ask again for the new origin.
  
  
  — — —
  
  HOW IT WORKS
  
  1. Ask. Type a task in plain English. Attach an image or audio file if useful (up to 10 MB each).
  2. See. The agent reads the page — structure, text, network activity — and plans the next step.
  3. Act. It clicks, types, scrolls, navigates, or uses the page's own WebMCP tools, always under your approval.
  
  The agent keeps going until the task is done or it needs you.
  
  
  — — —
  
  WHAT IT'S GOOD FOR
  
  • Fill long forms in seconds — job applications, insurance claims, support tickets.
  • Scrape structured data from paginated search results or authenticated pages.
  • Run manual QA on autopilot — replay user journeys, validate error states, confirm a new flow still works after a deploy.
  • Aggregate information across multiple sites without a scraping farm.
  • Compare prices across retailers, fill out the winner's checkout, and pause for your approval before payment.
  • Debug and inspect live pages — read console messages, monitor network traffic, run ad-hoc scripts (with your approval).
  
  
  — — —
  
  FOR DEVELOPERS AND PAGE AUTHORS
  
  Auto Browser speaks WebMCP — the open web specification for pages to describe themselves to AI agents. If you ship a web app, you can expose domain-specific tools to the agent in a few lines of code or a single HTML attribute. The agent prefers your tools over generic DOM automation, so every workflow that uses a WebMCP-tagged page is more reliable.
  
  Learn more: https://github.com/webmachinelearning/webmcp
  
  
  — — —
  
  REQUIREMENTS
  
  • Chrome 126 or newer for the cloud providers (Google Gemini, OpenRouter, or a local endpoint).
  • For the on-device Gemini Nano option: Chrome 138 or newer on macOS 13+, Windows 10+, Linux, or ChromeOS, with at least 16 GB RAM and 22 GB of free disk space.
  
  
  A NOTE ON PERMISSIONS
  
  Auto Browser uses Chrome's debugger API — this is how the agent reads and acts on pages. It is the reason Chrome shows a "being debugged" banner during a task. This is expected and by design; the banner is your visible signal that automation is in progress. All data handling is described in the Privacy Policy.

  Auto Browser is an AI agent that uses your browser for you.
  
  It sits in your Chrome side panel. You tell it what you want in plain English ("find me a flight from NYC to Tokyo next Thursday", "compare these soccer boots and pick one", "pull the top 5 rows from this table"), and it goes and does it. There are no scripts to write, no servers to set up, no per-site integrations.
  
  It works on any website, it lets you pick which AI runs it, and it can run fully on-device if you want zero data leaving your machine.
  
  
  HOW IT WORKS
  
  You type a task. The agent reads the page through its accessibility tree (so it understands structure, not pixels) and figures out what to do. Then it clicks, types, scrolls, and navigates on your behalf. It keeps going until the job is done or it needs you to step in.
  
  Every action that changes something on the page asks for your approval first. Reading the page never does.
  
  
  BRING YOUR OWN AI
  
  Most agent tools lock you into one model. Auto Browser doesn't. You pick from four providers in a dropdown, and you can switch mid-conversation without losing context.
  
  1. Chrome Built-in AI (Gemini Nano). Free, on-device, private. Runs locally inside Chrome with no API key, no account, no network calls. If you pick this, zero data leaves your browser.
  
  2. Google Gemini. Direct API access to Google's latest Gemini models. Your key, your quota.
  
  3. OpenRouter. One key gets you every model the OpenRouter catalog like DeepSeek. Image and audio support is detected per model.
  
  4. Local LLM. Point it at Ollama, LM Studio, or any OpenAI-compatible endpoint running on your machine. Good for air-gapped setups and custom fine-tunes.
  
  Start a task on the on-device model, hit something hard, flip to Claude, finish, flip back. Your conversation comes with you.
  
  
  WEBMCP NATIVE
  
  Auto Browser speaks WebMCP, an open web spec that lets pages describe themselves to AI agents. When a site exposes WebMCP tools, the agent uses those tools directly instead of clicking around the DOM. It is faster, cleaner, and less likely to break when the site changes.
  
  WebMCP is an open spec, not something we own. We just think it is the right way to build the web for agents.
  
  
  PRIVACY AND SAFETY
  
  An agent that can click anything is dangerous if you do not put guardrails on it. We put guardrails on it.
  
  There are no analytics in this extension. No telemetry. No usage tracking. Your API keys live in Chrome's storage on your machine.
  
  Passwords and secret tokens get redacted before they can reach the model. The agent can tell that a field has a value in it, but it never sees the value.
  
  Banking, government, and healthcare sites are blocked by default and you cannot override that inline. If a page redirects mid-task (a login loop, a phishing hop), any approvals you already gave stop applying and the agent has to ask again for the new origin.
  
  Running arbitrary JavaScript is the most powerful thing the agent can do, so it is the most gated. Even broad approvals never silently cover script execution.
  
  If you want absolute privacy, run Gemini Nano. Nothing leaves Chrome.
  
  
  WHAT PEOPLE ACTUALLY USE IT FOR
  
  Comparing products across a store and getting a recommendation. Pulling structured data out of nasty HTML tables. Summarising long articles, papers, and release notes. Reading what someone is posting on X without scrolling for twenty minutes. Booking restaurant reservations and movie tickets. Filling out tedious multi-step forms. Anything you keep putting off because it is boring.
  
  The point of a generic, app-agnostic agent is that one install handles radically different jobs.
  
  
  WHO IT IS FOR
  
  People who live in their browser. Researchers, analysts, founders, builders, operators, anyone who treats "lots of open tabs" as a permanent condition. If you spend hours a day inside Chrome, this is for you.
  
  
  GETTING STARTED
  
  Install the extension. Open the side panel on any site. Pick a provider (or stay on the free on-device one). Type what you want done. That is the whole onboarding.
  
  
  Auto Browser is closed-source software built by an independent team. Privacy policy and terms are linked from the listing.

• May 12, 2026
  short_description

  Generic AI orchestrator — connects your LLM to any web page via WebMCP

  AI agent that actually uses your browser. WebMCP-native, privacy-first, works with any AI providers.

• May 12, 2026
  permissions

  storage, unlimitedStorage, activeTab, tabs, sidePanel, scripting, debugger

  storage, unlimitedStorage, activeTab, tabs, sidePanel, scripting, debugger, offscreen

Permissions & access

Permissions

storageunlimitedStoragetabssidePanelscriptingdebuggeroffscreen

Host access

<all_urls>

Screenshots