Alibaba Releases Page Agent: In-Page AI for Web Automation

Alibaba has introduced Page Agent, an open-source, JavaScript-based library designed to bring agentic capabilities directly into web applications. Unlike traditional browser automation tools such as Playwright, Puppeteer, or Selenium, which operate from an external process, Page Agent functions from within the webpage itself. By running as client-side JavaScript, the agent acts as a real user, inheriting existing authentication, cookies, and session data without requiring a separate backend or headless browser setup.

DOM Dehydration and Natural Language Control

The core innovation behind Page Agent is a technique called DOM dehydration. Modern web pages often contain thousands of nodes, making them too complex and costly to process through large language models in their raw HTML form. To solve this, the agent scans the Document Object Model to identify interactive elements like buttons, links, and input fields. It then converts this data into a FlatDomTree, a clean, text-based map that strips away redundant markup.
This compact representation allows the agent to interpret and execute natural-language commands with high precision. Because the system is model-agnostic, developers can connect any OpenAI-compatible endpoint to the agent. Since the model only processes text rather than visual screenshots, the system remains efficient and cost-effective, relying on structured data to navigate interfaces.

Implementation and Developer Control

Page Agent is built with a TypeScript-first approach and is released under the MIT license. The library is structured as a monorepo, with the core logic residing in the core package and DOM extraction handled by the page-controller. Developers can integrate the agent into their own applications with minimal code, enabling features such as AI copilots, automated form filling, and voice-controlled accessibility via the Web Speech API.
To maintain security and operational integrity, the library provides several developer controls. Operation allowlists can restrict which actions the agent is permitted to perform, while data masking can prevent sensitive information, such as passwords, from being exposed to the model. Furthermore, developers can inject custom knowledge to ensure the agent adheres to specific domain rules.

Scope and Security Considerations

While Page Agent offers a powerful way to add agentic behavior to web apps, it is designed for specific use cases. It is best suited for copilots and internal tools where the developer has control over the code. Because it operates within a single-page scope, it cannot autonomously navigate across multiple tabs or windows without additional tools, such as the optional Chrome extension or a Beta MCP server.
Developers should also note that while prompt-level safety can guide the agent, it is not a substitute for robust security. For sensitive or destructive actions, such as submitting payments, the team recommends maintaining server-side validation. By keeping these critical checks on the backend, developers can leverage the convenience of natural language automation while ensuring that the application remains secure and compliant with its original UI validation rules.