Skip the headless browser: an AI agent that lives in the DOM

What it does

Page Agent is a client-side JavaScript library that lets an LLM operate a web interface through natural language commands. You load it into the page—via script tag or module—and it translates prompts into DOM actions like clicking and typing. The project treats the webpage itself as the runtime environment, not a headless browser or external automation server.

The interesting bit

Rather than burning tokens on screenshot-based vision models, Page Agent reads the DOM as structured text and feeds that to your own LLM. It borrows its DOM-processing logic from browser-use, but rearchitects the idea to live entirely inside the browser tab, which keeps the heavy infrastructure out of the equation.

Key highlights

• Runs fully in-page; no headless browser, browser extension, or backend automation required for single-page tasks.
• Text-only DOM manipulation means you can use cheaper, non-multimodal LLMs without special permissions.
• Optional Chrome extension and a beta MCP server extend it to multi-page workflows or external agent control.
• Alibaba ships a free demo LLM via CDN for quick evaluation, though production use requires your own API key and model.
• DOM processing and prompts are derived from the browser-use project, so it is not reinventing the wheel on parsing logic.

Caveats

• The MCP server is explicitly labeled beta, and multi-page support depends on an optional extension rather than the core library.
• The free evaluation LLM is restricted to technical testing and carries its own terms-of-use baggage.
• By design, it is client-side only; the README bluntly states it is not for server-side automation.

Verdict

A good fit for frontend developers who want to add an AI copilot or voice-controlled accessibility layer to an existing web app without rewriting the backend. Look elsewhere if you need server-side scraping or cross-tab orchestration that is production-ready out of the box.