AI-Powered Browser Assistant β Control any web interface with natural language, with a built-in Skill system for platform-specific precision automation.
π English | δΈζ
Page Steer is a browser extension + JS library that lets AI directly operate web pages. Instead of taking screenshots and guessing, it reads the live DOM, understands the page structure, and executes precise actions.
What makes it different:
- Skill System β Platform-specific recipes (like Twitter/X, Google Trends) that bypass DOM guesswork. Skills use stable selectors (data-testid, aria-label) for reliable, token-efficient automation.
- No screenshots, no vision models β Pure text-based DOM understanding. Works with any LLM.
- Three ways, one engine β Chrome extension (sidepanel chat) for everyday users, MCP server for Cursor/Claude Desktop, or embed as JS for your own app.
- Learns from use β AI records successful runs and replays them as reusable skills. Import/export to share.
Skills are platform-specific automation recipes that give the AI precise, reliable actions instead of relying on DOM parsing every time.
βββββββββββββββββββββββββββββββββββββββββββββββ
β User: "Post a tweet saying hello world" β
β β
β Agent detects: x.com β twitter skill β
β Agent calls: twitter_post_tweet({ β
β text: "hello world" β
β }) β
β β
β Skill executes via data-testid selectors β
β β Reliable, fast, no DOM guessing β
βββββββββββββββββββββββββββββββββββββββββββββββ
| Source | Description |
|---|---|
| Built-in | Ships with the extension (Twitter/X, Google Trends) |
| Learned | AI records successful interactions and saves as reusable recipes |
| Community | Import/export skills, share via the project repo |
| Skill | Actions |
|---|---|
| Twitter/X | get_timeline, post_tweet, like, retweet, reply, search, get_profile, get_tweet_detail |
| Google Trends | search_explore, compare_terms, get_page_data, set_location, set_time_period, set_category, get_trending_now, get_trending_detail |
Skills target stable selectors (React data-testid, aria-label) for reliable, CSP-safe automation. More platforms coming β or write your own.
βββββββββββββββββββββββββββββββββββββββββββββββββββββββ
β Chrome Extension β
β βββββββββββββ ββββββββββββ ββββββββββββββββββββ β
β β Sidepanel β β MCP β β Background β β
β β (Chat UI) β β Server β β (Skill Registry β β
β β + Slash β β (Native β β + Executor) β β
β β Menu β β Host) β β β β
β βββββββ¬ββββββ ββββββ¬ββββββ ββββββββββ¬ββββββββββ β
β β β β β
β ββββββββββββββββ΄βββββββββββββββββββ β
β β β
β PageSteerCore (Agent Loop) β
β ββββββββββ΄βββββββββ β
β β Skill Tools β β Auto-injected β
β β + DOM Tools β based on URL β
β ββββββββββ¬βββββββββ β
β β β
β PageController (DOM Ops) β
βββββββββββββββββββββββββΌβββββββββββββββββββββββββββββββ
β
Target Tab
| Package | Purpose |
|---|---|
@page-steer/page-controller |
DOM extraction + visual feedback |
@page-steer/llms |
LLM client with reflection-before-action |
@page-steer/core |
Headless agent core |
@page-steer/ui |
Panel UI + i18n |
page-steer |
Main entry with UI (JS library) |
packages/mcp |
MCP server for external AI clients |
packages/extension |
Chrome extension (WXT + React) |
- Clone and build:
git clone https://github.com/cookiebody/PageSteer.git
cd page-steer
npm install
npm run build:ext-
Load in Chrome:
- Go to
chrome://extensions - Enable Developer mode
- Click "Load unpacked" β select
packages/extension/.output/chrome-mv3
- Go to
-
Open sidepanel, configure your LLM endpoint, and start chatting.
cd packages/mcp
node src/install.jsThen add to your MCP client config. See MCP docs.
npm install page-steerimport { PageSteer } from 'page-steer'
const agent = new PageSteer({
model: 'your-model',
baseURL: 'https://your-api-endpoint/v1',
apiKey: 'YOUR_API_KEY',
})
await agent.execute('Fill in the contact form with test data')- Browser Automation β "Check my timeline" / "Post a tweet" / "Search AI agents"
- SaaS AI Copilot β Embed in your product, ship AI copilot in lines of code
- Smart Form Filling β Turn 20-click workflows into one sentence
- Multi-tab Operations β Agent works across browser tabs via the extension
- MCP Integration β Let Cursor/Claude Desktop control your browser
npm run dev:ext # Extension dev mode
npm run dev:demo # Demo dev server
npm run build # Build all packages
npm run build:ext # Build + zip extension
npm run typecheck # Type check
npm run test # Run testsWe welcome contributions! See CONTRIBUTING.md for guidelines.
DOM processing components derived from browser-use (MIT License, Copyright (c) 2024 Gregor Zunic).
β Star this repo if you find Page Steer useful!