An AI agent that actually uses your browser.

Auto Browser runs an AI agent in Chrome's side panel. Describe a task in plain English, and it reads the page and clicks and types for you, with your tabs and logins. By default it asks before each action.

Install free Watch how it works
  • No setup for each site
  • Your choice of AI, including local models
  • Free, with no account
webmcp-sports.com
  • Chrome extension, Manifest V3
  • WebMCP support
  • No telemetry
  • Works with local models
  • API keys stay in your browser
How it works

Three steps, repeated.

The agent reads the page, picks the next action and takes it. It repeats until the task is done, it needs your input, or it reaches the step limit (50 by default).

  1. 01

    Ask

    Type the task in plain English. You can attach images or audio when your model accepts them.

    “Find me a flight from NYC to Tokyo next Thursday, one-way.”
  2. 02

    See

    The agent builds an outline of the page's buttons, links, fields and text. It takes a screenshot only when it needs to check something visually.

    Snapshot taken → 47 elements, search form detected
  3. 03

    Act

    It clicks, types, scrolls, opens pages and tabs, or calls the page's own WebMCP tools. In the default mode it asks before each change.

    Fill form → submit → read results
Why Auto Browser

Four commitments.

No setup per site

It works on the pages you already use, signed in as you. Sites don't need to support it, and there's nothing to connect.

Your choice of AI

Run tasks with OpenRouter (DeepSeek V4.1 Flash by default), Google Gemini with your own key, or a local model through Ollama or LM Studio. Gemini Nano is there for on-device chat.

Learn more

Private by default

The extension has no analytics, no telemetry and no servers. API keys stay in your browser, and what you type into form fields stays out of what the model reads. With a local model, nothing goes to a cloud provider.

Learn more

Asks before it acts

By default, every click, keystroke and navigation waits for your approval, and scripts need an approval of their own. Banking, government and healthcare sites on a built-in list are blocked.

Learn more
Use cases

Six tasks, start to finish.

Each card is a recording of the agent doing the task. Tap one to watch it.

Shop

Compare products and recommend one

The agent searches the store, reads the specs and reviews for each option, and gives you a side-by-side summary with a recommendation.

“Compare all the soccer boots and suggest which one to buy.”
Extract

Pull rows out of a table

The agent writes a short script that filters and sorts the table, runs it once you approve, and returns only the rows you asked for.

“Give me the top 5 models cheaper than $2.”
Browse

Summarise what someone's posting

Open an X profile and ask what the person is posting about. The agent scrolls the timeline, reads the posts and writes a short brief.

“What is Elon Musk talking about?”
Read

Summarise long articles

Open a long post, paper or release note and ask for a summary. The agent reads the whole page and returns a summary with the main points.

“Summarise this article and give me the key insights.”
Reserve

Make a restaurant reservation

It finds the booking form, fills in the details you gave it and asks you for anything missing.

“Make a reservation for 2 next Friday, 7pm.”
Tickets

Book a movie ticket for Friday

It reads the showtimes, asks which film you want, and goes through seat selection and checkout, waiting for your approval before the final step.

“Let's see what movies are on and get a ticket for this Friday.”
AI providers

Pick the AI, and change it when you like.

OpenRouter, Google Gemini and a local model can all run tasks on the page, and you can switch between them mid-conversation. Gemini Nano runs on your device for chat.

01

Chrome Built-in AI

Free, on-device, for chat.

Gemini Nano runs inside Chrome on your computer, with no API key or account. Use it to chat and to ask about images or audio. Its context window is too small to drive web pages, so tasks on a page need one of the other three providers.

  • No API key
  • Runs offline after a 3 to 5 GB download
  • Chat only, no page actions
Best for Private chat · Offline use
02

Google Gemini

Google's models, called directly.

Requests go from your browser straight to Google's Gemini API, billed to your own key. The default model is Gemini 3.8 Flash.

  • Direct Google API
  • Latest Gemini models
  • Your key, your quota
Best for Complex reasoning · Long context
03

OpenRouter

Many models, one key.

Use any model in the OpenRouter catalog with one key. The default is DeepSeek V4.1 Flash, and Claude, GPT and Llama are a dropdown away. Auto Browser checks which models accept images and audio.

  • DeepSeek · Claude · GPT · more
  • One API key
  • Detects image and audio support
Best for Everyday tasks · Comparing models
04

Local LLM

Runs on hardware you control.

Point it at Ollama, LM Studio or any OpenAI-compatible endpoint. Nothing goes to a cloud provider, and your own fine-tuned models work too.

  • Ollama / LM Studio
  • OpenAI-compatible
  • No cloud provider
Best for Private tasks · Custom models

Switch mid-conversation.

Move between OpenRouter, Google Gemini and a local model, and the conversation and tool results carry over. Start on the default model and switch for a harder step.

Auto Browser settings: four LLM provider tabs (Gemini, OpenRouter, Built-in AI, Local) with a backend chooser (Ollama, LM Studio, Custom) and base URL field.
Under the hood

The parts that do the work.

Six pieces that decide how well it handles a task. The engineering posts on the blog cover each one in detail.

Reads the whole page

It builds an outline of the page that includes same-origin iframes, open shadow roots and content below the fold, and leaves out elements hidden behind a modal. It takes a screenshot only when a task needs a visual check.

Uses the page's own tools

When a page offers WebMCP tools, the agent is told to use them before clicking around. Each call goes through the same permission check as a click.

Asks before it acts

In the default mode, clicks, typing and navigation wait for your approval, and custom scripts need an approval of their own. Sites on a built-in list of banking, government and healthcare domains are blocked outright.

Recovers, or tells you why it stopped

If an element has changed since the agent last looked, it gets a fresh view of the page in the same step. When a task stalls, it changes plan once, then stops and says why.

Works across tabs

Within one task it can open, switch between and close tabs, working on one at a time. It sees only the tabs in its own Chrome tab group, and when the task ends it ungroups them without closing them.

Resumes after a restart

Chrome can stop an extension's background worker mid-task. Auto Browser then tries to resume the task, up to twice within 10 minutes, and tells you if it can't. It never resends an action whose result it didn't see.

92.7% WebVoyager-mini pass rate

1.3¢ Model cost per task

DeepSeek V4.1 Flash on Auto Browser 1.6.0. 103 of WebVoyager's 643 tasks, mean of two runs, self-reported. Read the results

Read the engineering posts
For developers

Your site, agent-ready.

Auto Browser supports WebMCP, a draft web standard that lets a page give AI agents its own tools. Register tools or mark up your forms, and the agent calls them in place of clicking through the page.

  • An open draft standard. The W3C Web Machine Learning Community Group writes WebMCP in the open, and any agent can implement it.
  • Usually more reliable than clicking. A registered tool states what it does and which inputs it takes, so the agent doesn't have to work it out from the page layout.
  • Nothing to install. On Chrome without native WebMCP, Auto Browser provides document.modelContext itself, following the current draft. You don't ship an SDK.
  • Same permission rules as a click. Every page tool call goes through the user's permission settings. A page can label its tool as read-only, but that label never skips the check.
Read the WebMCP spec
Open spec

WebMCP

A draft standard for web pages to offer tools to AI agents, written by the W3C Web Machine Learning Community Group.

github.com/webmachinelearning/webmcp
Privacy & safety

It asks before it acts.

An agent that clicks in your signed-in browser can buy, send or delete things. These checks limit what it can do without you, starting with the permission mode you pick.

  • Asks first by default

    In Ask mode, the agent asks before each click, keystroke and navigation. Reading the page and scrolling don't interrupt you. You can switch to auto-approve for one task, or for all actions until you turn it off.

  • Scripts need their own approval

    Running custom JavaScript on a page is the most powerful thing the agent can do, and auto-approve never covers it. You can allow scripts on one site for the rest of a task, or turn the prompt off in Settings.

  • Sensitive sites are off limits

    On a built-in list of about 30 banking, payment, crypto, government, healthcare and password-manager sites, the agent can't click, type, run scripts or open tabs, whatever mode you're in. It can still read a listed page you're on.

  • Approvals stay with their site

    In Ask mode, an approval applies to the site you gave it on. If the page has moved to another site before the agent acts, it asks again.

  • Typed values stay private

    What's typed into form fields, passwords included, stays out of the page outline the model reads. The model can see that a field is filled in, but not what it contains.

  • Only its own tabs

    The agent works in the tab you started from and in tabs it opens, all kept in one Chrome tab group. It can't see or act on your other tabs.

  • A local option

    With a local model through Ollama or LM Studio, the agent runs tasks without sending pages or prompts to a cloud provider.

app.netlify.com — Project configuration
  1. go to project overview
  2. Check current page for navigation to project overview

  3. take_snapshot

    Done
  4. Click Projects link in sidebar to navigate to project overview

  5. click

    Running Done Denied
  6. Allow Auto Browser to click on app.netlify.com?

  7. Navigation complete. You're on the project overview now.

  8. Navigation complete. I won't prompt again for clicks on this domain this turn.

  9. Action skipped. Tell me what you'd like to do instead.

Try it: choose an option above. You approved this single click. The next write will still ask. Mutating actions on this domain won't prompt again until the turn ends. The agent stops and waits for your next instruction.
Read the full privacy policy
FAQ

Questions, answered.

  • Is Auto Browser free?
    Yes, the extension is free. Gemini Nano and local models cost nothing to run. OpenRouter and Google Gemini bill your own API key per request. On OpenRouter's default model, our benchmark tasks averaged 1.3 cents each.
  • How well does it work?
    On WebVoyager-mini, 103 tasks from the WebVoyager benchmark run on live websites, it passed 92.7% of graded tasks with DeepSeek V4.1 Flash. That is the mean of two runs, self-reported, on Auto Browser 1.6.0. The post covers the method and its limits. Read the benchmark →
  • Which AI provider should I pick?
    For most tasks, use OpenRouter with its default model, DeepSeek V4.1 Flash. In our benchmark it scored within noise of Claude Sonnet 5.5, which cost 71 times as much. If pages must not go to a cloud provider, use a local model through Ollama or LM Studio. Gemini Nano also runs on your device, but only for chat: its context window is too small to drive web pages.
  • Do I need an API key?
    Only for OpenRouter and Google Gemini. Gemini Nano and local models need none. Keys are stored in your browser and sent only to their provider. Auto Browser has no servers.
  • Does my browsing data leave the browser?
    With a cloud provider, the agent sends that provider what the model needs for the task: your messages, an outline and the text of the page, and a screenshot when it takes one. With a local model, nothing goes to a cloud provider. The extension sends no telemetry.
  • Which sites does it work on?
    Most sites. The agent acts on the rendered page through the Chrome DevTools Protocol, so it sees what you see, including pages you're signed in to. It won't click, type or run scripts on a built-in list of about 30 banking, payment, government, healthcare and password-manager sites.
  • Can it work across several tabs?
    Yes. Within one task the agent can open tabs, switch between them and close the ones it opened, working on one tab at a time. It keeps them in a Chrome tab group called Auto Browser and can't touch your other tabs. When the task ends, the group is removed and the tabs stay open. Multi-tab tasks don't run in incognito or on a message with an attachment.
  • Can I swap providers mid-conversation?
    Yes, between OpenRouter, Google Gemini and a local model. Pick another in Settings, and the conversation and tool results carry over. You stay signed in to the site because the agent uses your own Chrome session. Gemini Nano keeps a separate chat.
  • Why does Chrome say the browser is being debugged?
    Auto Browser clicks and types through the Chrome DevTools Protocol, and Chrome shows that banner whenever an extension uses it.
  • What are the install requirements?
    A recent version of Chrome on Windows, macOS, Linux or ChromeOS. The on-device Gemini Nano option needs Chrome 138 or newer, macOS 13+, Windows 10+, Linux or ChromeOS Plus, 16 GB of RAM and 22 GB of free disk.
  • How do I uninstall?
    Open chrome://extensions, find Auto Browser and click Remove. Chrome deletes the extension's data, including API keys, conversation history and settings, and there is no account to cancel. The Gemini Nano model belongs to Chrome and stays installed. Anything already sent to a cloud provider falls under that provider's own policy.
  • How is this different from ChatGPT agent or Claude in Chrome?
    Auto Browser works in your own Chrome, with your tabs and logins, and with the model you choose, including a local one. It needs no account, and when a page offers WebMCP tools it uses them under the same permission checks as a click. Agents from AI companies work with that company's own models and need an account with it.
  • Who builds Auto Browser?
    Auto Browser is an experimental project by Vin Lim, who also writes the engineering posts on this blog. About Vin Lim →

Stop doing it by hand.

Install it free, pick an AI provider, and describe the task.

Install free Read the docs

Free, with no account to create.