Page Agent

What is Page Agent?

Page Agent is a GUI agent developed by Alibaba that runs entirely inside your web page. The pitch is delightfully simple: add a single script tag to any page and that page now has its own AI agent, one that can understand the page structure and act on it when you give it a plain-language instruction like "click the login button" or "fill in this form".

The clever part is how it works. Instead of taking screenshots and asking a vision model what it sees, Page Agent works with the text of your DOM directly. That makes it lightweight, fast, and compatible with any modern browser, and it means the agent can run on plain text-based LLMs rather than expensive multimodal ones. Most mainstream models work, including models you host locally.

The project is MIT-licensed and open source, with the code on GitHub and a project page on GitHub Pages. Because it is just JavaScript, you can embed it in an existing product without rewriting your backend, which is why it has become a popular way to bolt an AI copilot onto SaaS apps.

Key Features

• One-line integration: a single script tag brings the agent to any page
• Text-based DOM manipulation, no screenshots or multimodal LLMs needed
• Works with OpenAI-compatible APIs and locally deployed models (Ollama, vLLM, llama.cpp)
• Bring your own LLM: pick model, endpoint, and API key in the config
• Runs fully in-page via JavaScript, no Python backend or headless browser
• Chrome extension extends the agent across multiple tabs for multi-page tasks
• MCP Server (Beta) lets external MCP clients drive the browser through the agent
• Free testing LLM API bundled with the CDN demo build

How to get started

The fastest way to try Page Agent is the one-line CDN integration. The demo build includes a free testing LLM API and creates a demo agent automatically, so you can play with it in minutes on any page. Add ?autoInit=false to the script URL if you would rather instantiate the agent yourself with new window.PageAgent(...).

For production, install the npm package and configure your own LLM provider. The config object takes the model name, an OpenAI-compatible base URL, your API key, and a language setting. Because it speaks the OpenAI API dialect, it pairs with virtually any provider, from hosted APIs to models you run locally.

• Add the CDN script tag to a test page, or run npm install page-agent
• Create an agent instance with your model, baseURL, apiKey, and language
• Try a simple command like clicking a button or filling a form
• Test with the bundled demo LLM before wiring up your own provider
• Move to the npm build and configure your production LLM endpoint

Use cases

• AI copilot for SaaS: give any web app an assistant without touching the backend
• Workflow automation: turn 20-click form filling into one sentence for ERP, CRM, and admin systems
• Accessibility: let users drive any web app through natural language or voice commands
• Multi-page tasks: coordinate actions across browser tabs with the Chrome extension
• MCP control: let AI coding agents and MCP clients browse pages and fill forms through the agent

Pricing and licensing

Page Agent itself is free. The project is MIT-licensed open source, so you can use it, modify it, and embed it in commercial products without paying a license fee. That makes it an attractive option compared to proprietary agent platforms that charge per task or per seat.

What you pay for, if anything, is the LLM behind the agent. You bring your own model and pay your provider's API rates, or run a local model for free on your own hardware. The CDN demo build even includes a free testing LLM API so you can evaluate the agent before wiring up a paid provider.

Pros and cons

Page Agent shines on simplicity and cost. One script tag beats installing extensions, standing up a backend, or renting a headless browser, and the text-based DOM approach keeps token usage low compared to screenshot-driven agents. The MIT license and bring-your-own-LLM design give you full control over both cost and data.

The trade-offs come from being browser-native. The agent only sees what the page's DOM exposes, so it can struggle with canvas-based apps, shadow DOM edge cases, or content rendered in iframes. It is also a young project, so the feature set, docs, and ecosystem are still growing compared to more established browser automation tools.

Alternatives

• Browser Use: open-source agent that drives a real browser for automation
• Stagehand: natural language browser automation framework built on Playwright
• Skyvern: AI agent for browser-based workflows with vision capabilities
• n8n: workflow automation platform with AI agent nodes for server-side flows
• Selenium or Playwright: classic code-first browser automation for developers