Page Agent is an MIT-licensed TypeScript library from Alibaba that embeds a natural-language GUI agent directly inside a web page, aimed at product teams, accessibility engineers, and automation developers who want to drive their own interfaces with plain-language instructions instead of hand-written scripts.

What it is

Page Agent describes itself as the GUI agent living in your webpage: one script gives any web page its own AI agent. It is distributed as the npm package page-agent, written in TypeScript, published under the MIT licence by the alibaba organisation, and filed in this registry under Data & Analytics / Data Extraction & Web Scraping. The agent runs as in-page JavaScript, performing text-based DOM manipulation rather than screenshot analysis, which removes the need for multi-modal LLMs and special browser permissions. It works with most mainstream models, including locally deployed ones, and ships an optional Chrome extension for multi-page tasks plus an MCP Server in Beta for control from outside the page.

The concrete problem it solves is external automation plumbing. Driving a web interface normally means a browser extension, a Python stack, or a headless browser, each carrying its own install, permission, and maintenance burden. Page Agent replaces that with a script tag or an import: the page itself hosts the agent, so there is no separate driver to keep in step with the interface, no backend rewrite to expose a copilot, and no screenshot pipeline to operate.

Key capabilities

• One-line CDN integration that boots a working agent with a free demo LLM; append ?autoInit=false to load the script without auto-creating the demo agent, then instantiate with new window.PageAgent(...).
• NPM install path: npm install page-agent, then import { PageAgent } from 'page-agent' and construct an agent with model, baseURL, apiKey, and language options before calling await agent.execute('Click the login button').
• Text-based DOM manipulation with no screenshots, so no multi-modal LLM and no special browser permissions are required.
• Bring-your-own-LLM configuration, demonstrated against qwen3.5-plus through the OpenAI-compatible endpoint https://dashscope.aliyuncs.com/compatible-mode/v1.
• Optional Chrome extension (web store ID akldabonmimlicnjlflnapfeklbfemhj) that extends a single-page agent across browser tabs.
• MCP Server (Beta) so external agent clients can control the browser.
• Dual-language documentation, English and Chinese via docs/README-zh.md, with CI defined in .github/workflows/main-ci.yml and zod declared as a peer dependency excluded from the published bundle-size figure.

Who uses it and how

• SaaS product teams ship an AI copilot inside their own application in lines of code, with no backend rewrite.
• Organisations running ERP, CRM, and admin systems turn twenty-click workflows into a single sentence through smart form filling.
• Accessibility teams expose any web app to natural-language control, covering voice commands and screen readers.
• Teams that need cross-tab behaviour install the Chrome extension and let one agent follow a task across multiple pages.
• Agent-platform builders connect their MCP clients to a browser and let the outside agent drive pages through the Beta MCP Server.

Getting started

Add the one-line CDN script to a page to try Page Agent against the free testing LLM API, or install the package with npm install page-agent and pass your own model, baseURL, and apiKey to new PageAgent(...); full programmatic usage is documented at alibaba.github.io/page-agent/docs/introduction/overview.

How it compares

No list of paid products this project replaces is provided in the facts, and no comparable named tools appear either, so Page Agent stands alone in this registry. The README positions it against approaches rather than products: it removes the need for a browser extension, a Python stack, or a headless browser by moving the agent into the page itself.

When to use it — and when not to

Pick Page Agent when the interface you want to automate is your own web page and you can supply an LLM endpoint and API key; there is no database, object storage, or SMTP service to operate, since the agent runs client-side with the model you point it at. Do not treat the CDN demo as production: it uses a free testing LLM API and the README marks it for technical evaluation only, with separate terms. Weaknesses to weigh are that the MCP Server is still Beta, cross-page tasks depend on installing the optional Chrome extension, and any deployment sends page content to whichever model endpoint is configured, which rules it out where DOM data cannot leave the browser or the organisation.

<script
    src="https://cdn.jsdelivr.net/npm/[email protected]/dist/iife/page-agent.demo.js"
    crossorigin="anonymous"
></script>

<!-- China CDN mirror if you can't access jsDelivr -->
<!-- https://registry.npmmirror.com/page-agent/1.12.4/files/dist/iife/page-agent.demo.js -->

npm install page-agent

import { PageAgent } from 'page-agent'

const agent = new PageAgent({
    model: 'qwen3.5-plus',
    baseURL: 'https://dashscope.aliyuncs.com/compatible-mode/v1',
    apiKey: 'YOUR_API_KEY',
    language: 'en-US',
})

await agent.execute('Click the login button')

DOM processing components and prompt are derived from browser-use:

Browser Use <https://github.com/browser-use/browser-use>
Copyright (c) 2024 Gregor Zunic
Licensed under the MIT License

We gratefully acknowledge the browser-use project and its contributors for their
excellent work on web automation and DOM interaction patterns that helped make
this project possible.