Docs

Documentation

Install Auto Browser, pick an AI, and get the most out of the side panel. A quick reference for everyday use and WebMCP integration.

Last updated April 2026

Auto Browser is a Chrome extension that puts an AI agent in your side panel. Describe a task in plain English; the agent reads the page and acts on your behalf — clicking, typing, navigating, and filling forms under your supervision.

This page covers installation, day-to-day use, choosing an AI provider, and the WebMCP interface for pages that want to expose their own tools.

Install

  1. Install Auto Browser from the Chrome Web Store.
  2. Click the toolbar icon to open the side panel.
  3. Open Settings and pick an AI provider (see below).
  4. Type a task. Review each action the agent wants to take. Approve or deny.

Requirements

  • Chrome 126 or newer for the cloud providers.
  • For the on-device Gemini Nano option: Chrome 138+, macOS 13+ / Windows 10+ / Linux / ChromeOS, 16 GB RAM, and 22 GB free disk.

Using the side panel

The side panel is a chat interface. Ask for something; the agent works across the current tab.

  • Describe the task in plain English. “Find me the cheapest one-way flight from NYC to Tokyo next Thursday.” The more specific, the better.
  • Attach files (paperclip icon) — images and audio, up to 10 MB each. Useful for uploading a resume, a screenshot, or a voice note.
  • Switch providers mid-conversation. Your history carries over. Try Gemini Nano for a task, flip to Claude on OpenRouter if it gets tough, flip back.
  • Stop or steer anytime. The agent asks before each write. You can deny, rephrase, or say “stop” — it always yields.

Choosing an AI provider

Four options, picked from a single dropdown in Settings. Each has trade-offs:

Chrome Built-in AI (Gemini Nano)

Free. On-device. Private.

Runs locally inside Chrome. No API key, no network calls. Best for privacy-sensitive work and offline use. Requires a one-time download of the model (~3–5 GB).

Google Gemini

Frontier quality, direct API.

Sends prompts directly to Google with your API key. Best for complex reasoning where model quality matters most.

OpenRouter

Every model, one key.

Any model in the OpenRouter catalog — Claude, GPT, Llama, Mistral, and others. Image and audio support is auto-detected per model. Best for experimenting and comparing.

Local LLM (Ollama-compatible)

Your machine. Your rules.

Point at any OpenAI-compatible endpoint on your network, including Ollama and LM Studio. Best for air-gapped work, custom fine-tunes, and zero cloud dependency.

Rule of thumb: start with Gemini Nano if you can run it. Move to OpenRouter or Gemini when a task is too hard. Switch to a local model when a task needs to stay on your machine.

Safety

Auto Browser is built to ask, not assume.

  • Every write action prompts. Clicks, typing, navigation — the agent requests approval before it fires. Read-only actions (looking at the page) don’t need your confirmation.
  • Banking, government, and healthcare sites are blocked by default. The agent is disabled entirely on those origins; there’s no way to override it inline.
  • Arbitrary JavaScript execution always asks. Even if you’ve given broad approval for a site, script execution prompts every time.
  • Passwords and secret tokens are redacted. The agent can tell that a field has a value, but never sees the value itself.
  • If a page redirects mid-action, permission resets. Login loops and phishing hops can’t sneak past an earlier approval.

For the full policy and what data (if any) reaches your chosen AI provider, see Privacy.

WebMCP — for page authors

Auto Browser supports WebMCP, the open spec for pages to describe themselves to AI agents. If you tag your page with WebMCP, the agent uses your tools instead of guessing its way through the DOM — more reliable, more predictable, more explicit.

The spec, API reference, and examples live at the official repo: github.com/webmachinelearning/webmcp.

Troubleshooting

The side panel doesn’t open. Make sure you’re on Chrome 126 or newer. Click the toolbar icon (or right-click any page → Auto Browser).

Gemini Nano says “downloading” forever. You need Chrome 138+, 16 GB RAM, 22 GB free disk, and a compatible OS. Check chrome://on-device-internals for model status.

Actions get blocked on a site. Banking, government, and healthcare origins are blocked by design. If a site you want the agent to use is being blocked incorrectly, let us know via the Chrome Web Store support channel.

The agent gets stuck in a loop. Stop the turn, rephrase the task more narrowly, or switch to a stronger model.

More

Occasional longer-form writing lives on the blog.