page-agent
An open-source web-embedded AI agent component released by Alibaba. It automates natural-language-driven page interactions through DOM structure parsing rather than screenshots, supports over 60 mainstream large language models, is licensed under MIT, and has zero server-side dependencies.
Timeline (last 90 days)
Alibaba has open-sourced page-agent, a graphical interface agent based entirely on web technology that runs embedded directly within website pages.
The current version of page-agent is V1.12.4, released under the MIT license.
page-agent has received more than 29,000 stars in the open-source community.
page-agent supports more than 60 mainstream large models, can be integrated with a single line of code, and has zero server-side dependencies.
page-agent cannot read purely visual content such as images, canvases, and 3D graphics, and the degree of page semantic markup directly determines the accuracy of AI comprehension.
page-agent offers three integration methods: CDN import, package manager (NPM) import, and offline operation via local runtimes such as Ollama or LM Studio.
Domestic models supported by page-agent include Tongyi Qianwen, DeepSeek, Zhipu GLM, Kimi, Tencent Hunyuan, and MiniMax; overseas models include OpenAI GPT, Claude, Gemini, and Grok.
page-agent provides four layers of controllability guarantees: operation allowlists/blocklists, data desensitization, knowledge injection, and human-in-the-loop confirmation.
page-agent supports click, text input, dropdown selection, page scrolling, form submission, focus switching, and same-origin single-layer iframes; JavaScript execution is an optional toggle that is off by default.
page-agent does not support hover, drag-and-drop, right-click menus, keyboard shortcuts, coordinate-based positioning, nested frames, cross-origin frames, drawing, or code editors requiring script control.
1 more timeline events
All Facts (9)
page-agent and browser-use have different positioning: the former is a page-embedded component that manages the current page and is aimed at website developers for enhancing user experience; the latter is an external tool that manages the entire browser and is aimed at crawler and agent developers for backend automation.
50%Unverifiedpage-agent supports more than 60 mainstream large models, can be integrated with a single line of code, and has zero server-side dependencies.
50%Unverifiedpage-agent cannot read purely visual content such as images, canvases, and 3D graphics, and the degree of page semantic markup directly determines the accuracy of AI comprehension.
50%Unverifiedpage-agent offers three integration methods: CDN import, package manager (NPM) import, and offline operation via local runtimes such as Ollama or LM Studio.
50%UnverifiedDomestic models supported by page-agent include Tongyi Qianwen, DeepSeek, Zhipu GLM, Kimi, Tencent Hunyuan, and MiniMax; overseas models include OpenAI GPT, Claude, Gemini, and Grok.
50%Unverifiedpage-agent provides four layers of controllability guarantees: operation allowlists/blocklists, data desensitization, knowledge injection, and human-in-the-loop confirmation.
50%Unverifiedpage-agent supports click, text input, dropdown selection, page scrolling, form submission, focus switching, and same-origin single-layer iframes; JavaScript execution is an optional toggle that is off by default.
50%Unverifiedpage-agent does not support hover, drag-and-drop, right-click menus, keyboard shortcuts, coordinate-based positioning, nested frames, cross-origin frames, drawing, or code editors requiring script control.
50%UnverifiedAlibaba has open-sourced page-agent, a graphical interface agent based entirely on web technology that runs embedded directly within website pages.
50%