
The GUI Agent Living in Your Webpage. Control web interfaces with natural language.

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: 'zh-CN',})
await agent.execute('帮我填写上报销单,金额 xxx,类别为 xxx')
开源地址:https://github.com/alibaba/page-agent
Currently, the mainstream path for automating browser operations with agents relies on:
Screenshots → Multimodal LLM recognizing elements → Executing actions.
This approach works, but it comes at a high cost:
Multimodal LLMs are expensive.
It requires running on servers or headless browsers, which is overly complex.
The open-source project introduced today — Page Agent — takes a different approach.
It embeds the agent directly into the web page:
Users give natural language commands, and the agent parses intent on the current page’s DOM to perform clicks, inputs, and other operations.
01 Project Introduction
Page Agent is a pure JavaScript in-page GUI agent.
Its purpose, as stated in the README.md, is clear:
Let AI live in web pages and control interfaces using natural language.
The project acknowledges browser-use, but the two have completely different positions:
browser-use is a server-side general-purpose browser automation engine.
Page Agent is its client-side version — lightweight and page-specific.
02 Project Highlights
1. Minimal Setup:
Traditional automation requires complex dependencies like Python environments and browser drivers.
Page Agent only needs a single script tag to get started:
2. Pure Text DOM Control:
Page Agent reads and manipulates the DOM structure directly:
Converts DOM to structured text for LLM processing.
The LLM plans the steps, and the built-in executor performs clicks, inputs, etc.
This is fast, low-cost, and highly accurate.
3. Open Model Ecosystem:
Page Agent uses a BYO LLM (Bring Your Own Large Language Model) architecture.
Supports all models compatible with OpenAI API format, including:
OpenAI, Claude Code, Qwen, Doubao, DeepSeek, etc.
Can also connect to local Ollama models for offline operation.
4. Cross-Page Capabilities: Chrome Extension + MCP:
(1) Chrome Extension: Enables control across tabs and multiple pages — useful for workflows requiring data transfer between pages A and B.
(2) MCP Server (Beta): Allows external agents (like Claude Code) to remotely control the browser — perfect for integrating browser control into larger agent workflows.
5. Privacy and Security:
Page Agent runs entirely within the browser; no data passes through third-party servers.
API keys are stored locally in the browser and not synchronized externally.
Already Followed
Follow
Replay Share Like
Close
Watch More
More
Exit Fullscreen
Fanxing AI Essay Collection Followed
Share Video
Duration: 00:18
0/0
00:00/00:18
Switch to Landscape Mode
Continue Playing
Progress bar, 0%
Play
00:00
00:18
00:18
Speed
Fullscreen
Speed Playback in Progress
0.5x 0.75x 1.0x 1.5x 2.0x
HD Smooth
Continue Watching
23.9K Stars! This open-source project turns web pages into AI agents — no screenshots, plugins, or backend required
Watch More
Republish
,
23.9K Stars! This open-source project turns web pages into AI agents — no screenshots, plugins, or backend required
Fanxing AI Essay Collection Followed
Share Like Read
Synced to Moments Write your comment
Video Details
03 Quick Start
1. Minimal Experience (evaluation only):
Load the page; an agent panel appears in the bottom right where you can input natural language commands.
2. NPM Installation:
(1) Install dependencies:
(2) Initialize agent:
(3) Execute natural language command:
04 Use Cases
1. SaaS Product AI Copilot:
Embed an AI assistant into your own product with just a few lines of code — no backend changes needed.
2. Smart Form Filling:
Compress 20-step ERP/CRM/backend operations into a single sentence.
3. Accessibility Enhancement:
Add natural language operation capabilities to any web app — ideal for voice commands and screen readers.
4. Cross-Tab Agent:
With the Chrome extension, agents can work across pages.
5. MCP Control:
Use MCP Server (Beta) to let external agent clients control the browser.
05 Conclusion
Page Agent’s value lies in:
Its focus on frontend nativity, natural interaction, and zero infrastructure — making AI truly integrate with web pages.
I’m Fanxing — follow me for more fun and useful open-source repos!
Thanks for reading~
See you next time!
It focuses on front-end native integration, natural interaction, and zero infrastructure, making AI truly embedded in web pages.
I'm Fanxing. Follow me for more fun and interesting open-source repositories!
Thank you for reading my article!
我们下期再见!