Breaking

Alibaba Open-Sources PageAgent, a JS AI Agent Running Inside the Web Page

July 4, 2026 at 20:15 EDT

AI Agents
Open Source
Software Dev & Coding

Alibaba has open-sourced PageAgent, a JavaScript AI agent that lives directly inside a web page and lets users control the entire interface with natural language, released under the MIT license. It can be dropped into a SaaS or web app with a single <script> tag or an npm install, letting users operate the screen by typing commands.

July 3, 2026 · Alibaba · Open Source

PageAgent: One Script Tag Turns Any Website Into an AI Copilot

Alibaba's MIT-licensed JavaScript agent "lives" inside the page — reading the live DOM as structured text and firing native browser events to drive the UI through plain natural language. No screenshots, no headless browser, no backend changes.

23.1k

GitHub stars (early July 2026)

Forks

Line of code to integrate (CDN script)

Tool price — pay only your own LLM

Inside-out, not outside-in

PageAgent flips the traditional automation model — instead of an external process poking at the browser, the agent runs client-side within the page itself.

Traditional automation

Selenium · Playwright · Puppeteer

Drives the browser from the outside
Vision models / screenshots
Needs headless browser, Python

PageAgent

Plain JavaScript inside the page

Reads the live DOM as text
Text-only LLM = lower cost
Inherits session & cookies, client-side

How it works · "DOM dehydration"

1 · SCAN

Find interactive elements; assign indices, roles, labels

→

2 · COMPRESS

Build lightweight text FlatDomTree for the LLM

→

3 · ACT

Dispatch native click, input, scroll events on the UI

Works with any OpenAI-compatible endpoint — Qwen · GPT · Claude · Gemini · DeepSeek · Grok, or local models via Ollama. Guarded by a human-in-the-loop approval panel, data masking and action allowlists.

What developers praise

Radical simplicity — one script tag
Cheaper: text models suffice
Structured DOM beats pixel-guessing
Client-side privacy; voice & accessibility
Automating ERP/CRM forms, legacy apps

What critics flag

Multi-tab needs the Chrome extension
CSP headers can block inline scripts
Prompt-based safety isn't a hard guarantee
Visual effects require WebGL2
Setup snags with local LLMs / extensions

Key specifications

License · MIT

Latest · v1.11.0

Tech · TypeScript ~82.6%

Install · CDN or npm

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.