Background

Every day, we work in browsers—searching for information, browsing web content, placing orders, organizing spreadsheets, filling out forms, and more.

On one hand, we open dozens or even hundreds of browser tabs, and switching and managing tabs has become an impossible task. Often, we need to start over and focus only on the tabs in front of us.

On the other hand, many tasks are repetitive, requiring us to execute them over and over again—filling out forms, organizing spreadsheets, solving CAPTCHAs, and so on. This process is both tedious and inefficient.

So, what if AI could do it? I previously built a tab management plugin that intelligently organizes cluttered tabs into groups. With the development of Tool Call and MCP, I realized that AI Agents could do more and more things. Soon, this plugin evolved into the current browser extension AIPex. AIPex supports using natural language to control the browser, and after optimizing context engineering, it can complete tasks quickly and accurately.

Through this article, I hope to answer these questions:

1. Why are AI browsers the future trend?
2. What are the implementation paths for AI browsers? What are their respective advantages and disadvantages?
3. Why are we confident that AIPex is the game changer for AI browser automation?

The AI Browser Revolution is Coming

Since the launch of ChatGPT, many teams have attempted AI browsers. The earliest was ChatGPT for Google, which could display AI responses alongside Google search results, instantly attracting millions of user registrations. Then, with the release of Sider and Monica, not only could they enhance Google search results, but they could also summon AI assistants on any page at any time, with specialized optimizations for video sites, chat PDFs, and image sites. AI could join in content generation, revision, and analysis at any time, greatly enhancing the user experience. I remain a loyal user of such plugins to this day. Sider alone has 6 million users on Chrome, making it arguably the number one AI browser plugin.

Later, with the emergence and widespread adoption of Tool Use and MCP, AI browsers could not only query and generate text and images in dialog boxes, but also call browser capabilities through tools to complete more complex tasks. From Browser Use proposing this concept last year, to the appearance of Claude for Chrome, Comet, and ChatGPT Atlas this year, various companies have released Agentic Browsers, with the biggest feature being the ability to automatically complete tasks for you. From Query to Action, user operations have shifted from passive browsing to active automation.

1. The fundamental bottleneck of traditional browsers: they only "display," they don't "complete"

The role of traditional browsers is:

Open web pages

Display information

Wait for people to click, copy, judge, navigate

But the real goal has never been "viewing web pages," but rather:

Finding the right product

Completing tax filing / booking tickets / filling forms

Comparing options and making decisions

Transforming information into results

What humans do in browsers is actually process execution, not "reading." The core upgrade of AI browsers is:

2. The content explosion of the modern web

The reality is:

• Pages are becoming increasingly complex (multi-step processes, pop-ups, CAPTCHAs, options)

• Information density is getting higher (price comparisons, terms, reviews, policies)

• Tasks are becoming more process-oriented (applications, registrations, comparisons, submissions)

While humans:

• Click slowly

• Have limited memory

• Make mistakes easily

• Are not good at repetitive processes

👉 It's not that humans aren't smart, but rather that humans aren't suited to be "web process engines"

AI browsers that autonomously execute tasks are essentially:

• Never tired

• Never miss steps

• Can execute in parallel

• Can continuously optimize paths

3. The key bottleneck of the information age has shifted from "obtaining information" to "executing decisions"

The problem in the past was:

• Too little information → We needed search engines

The problem now is:

• Too much information → Decision-making and execution costs are too high

Examples:

• Choosing the most cost-effective flight + baggage rules + change policies

• Sending customized emails to 20 suppliers and following up

• Completing a reimbursement / procurement process according to company policy

These aren't "just search and you're done," but mainly involve understanding goals, weighing constraints, and executing steps.

The traditional model is:

Human → Command → Software → Wait for feedback → Human operates again

The AI browser model:

Human → Set goals and boundaries
AI → Autonomous planning + execution + report results

4. Why browsers?

Computers were born as a medium between humans and information. Full computer use is indeed a trend for the more distant future, but browsers basically cover 90% of work and life systems, connecting to SaaS, government services, finance, and content platforms. Instead of waiting for each provider to offer APIs, we can directly complete tasks at the "human interface" layer. This is currently, in my opinion, the most realistic, least resistant, and most scalable path for AI automation implementation.

In one sentence, why we need AI browsers:

Implementation Principles for AI Browsers

The key to implementing AI browsers lies in how to efficiently understand web pages. Here are the following approaches:

1. DOM Tree — The most intuitive, yet also the most fragile approach

• Directly read document / HTML

• Serialize DOM nodes into text

• Hand over to LLM for understanding + generating actions

HTML / DOM → serialize → LLM → action

Playwright / Puppeteer also follow the DOM approach. They do a lot of dirty work in processing DOM, enabling them to get a relatively clean DOM tree representation. However, this approach has the following problems:

❌ DOM ≠ What users see on the interface

❌ div nested in div, DOM is not semantic expression leading to semantic loss

❌ LLM token explosion

2. Visual Tree / OCR (Visual Understanding Approach)

Treat web pages as "screenshots," use OCR + Vision Model to identify: buttons, text, input fields, then let AI click through coordinates

Screenshot → Vision Model → UI elements → click(x,y)

Currently, OpenAI also has a computer-use-agent (CUA) model that can generate actions based on screenshots and tasks. The advantage is that this approach is more universal, not dependent on the browser's representation of web pages, and can be extended to automation on any browser, any operating system. Although this solution is universal, it has high costs and latency. Currently, even ChatGPT Atlas does not use CUA for automation.

3. Accessibility Tree — AIPex's Approach

Principle (Key Point)

Browsers internally already have a "semantic tree for screen readers":

• role: button / textbox / link

• name: human-readable names

• state: disabled / checked / expanded

• hierarchy: real UI structure

DOM → Accessibility Tree → Semantic UI Graph → LLM

Why is it perfect for AI?

Product Forms of AI Browsers

Currently, there are the following product forms for browser automation. Let's analyze them one by one:

1. Agent Browsers

Agent browsers refer to standalone AI browser applications, such as Comet and ChatGPT Atlas. These products rebuild browsers from scratch, deeply integrating AI capabilities into the browser kernel.

Advantages:

• Deep Integration: AI capabilities are deeply integrated with the browser kernel, allowing more low-level control over browser behavior
• Unified Experience: All features are in one application, providing a more unified experience
• Performance Optimization: Can be specifically optimized for AI scenarios

Disadvantages:

• High Migration Cost: Users need to abandon their existing browsers and migrate bookmarks, extensions, passwords, and other data
• Ecosystem Fragmentation: Cannot use Chrome/Edge's rich extension ecosystem
• Learning Curve: Users need to adapt to new browser interfaces and operating habits
• High Development Cost: Requires building a browser from scratch, with extremely high development costs

Typical Representatives: Comet, ChatGPT Atlas, Dia

2. Extension/Plugin Approach

The extension/plugin approach refers to extension programs developed based on existing browsers (Chrome, Edge, etc.), such as AIPex. This approach adds AI automation capabilities on top of existing browsers.

Advantages:

• Zero Migration Cost: Retain all bookmarks, extensions, passwords, and history
• Plug and Play: Available immediately after installation, no need to change usage habits
• Ecosystem Compatibility: Can continue using Chrome/Edge's rich extension ecosystem
• High Development Efficiency: Based on mature browser APIs, relatively low development costs
• High User Acceptance: Users don't need to change their browser usage habits

Disadvantages:

• API Limitations: Limited by browser extension API capabilities
• Performance Constraints: Need to coordinate with other browser extensions and features, may be affected by performance

Typical Representatives: AIPex, Claude for Chrome

Path Comparison

From a practical implementation perspective, the extension/plugin approach is currently the most realistic, least resistant, and highest user acceptance path. Users don't need to abandon their established workflows and habits to gain AI automation capabilities. This is also the core reason why AIPex chose the extension path.

AIPex's Advantages

Product Advantages

1. No Migration Required

Unlike solutions like Comet and ChatGPT Atlas that require installing entirely new browsers, AIPex is a Chrome/Edge extension.

• Zero Migration Cost: Just install the extension and you're ready to use it
• Retain All Data: Bookmarks, extensions, passwords, history, cookies—all preserved
• No Habit Changes: Continue using familiar browser interfaces and operations
• Plug and Play: Available immediately after installation, no need to learn a new interface
• Ecosystem Compatible: Can continue using Chrome/Edge's rich extension ecosystem

As the AIPex GitHub repository says: "Your browser already works!" — Your browser is already great, we just make it smarter.

2. Open Source & Privacy Protection

For an AI Agent that can read and execute tasks, privacy and security are crucial. AIPex adopts the MIT open source license, completely transparent, auditable, and extensible:

• Fully Open Source: Code is completely public, anyone can review, contribute, and fork
• Privacy First: Your data never leaves your machine
• BYOK (Bring Your Own Key): Use your own API keys, completely control data flow

Compared to solutions like ChatGPT Atlas and Dia that require paid subscriptions and upload data to servers, AIPex has clear advantages in privacy and security.

3. Excellent Context Engineering

AIPex has made extensive optimizations in context engineering, which is the core technical advantage that enables it to complete tasks efficiently and accurately:

Accessibility Tree + Search Retrieval Mechanism:

• Uses semantically richer Accessibility Tree instead of traditional DOM
• Recalls relevant elements on-demand through semantic search, rather than passing the entire page
• Significantly reduces context length, improving response speed and accuracy

Intelligent Snapshot Deduplication:

• Only keeps the latest page snapshot for the same tab
• Reduces context complexity from O(n²) to O(n)
• 50 operations: from 1,275 snapshots down to 50 snapshots (96% token savings)

Search-based Element Retrieval:

When processing web content, AIPex does not use embedding-based RAG technology. Compared to code, web pages are constantly changing, and static embeddings are difficult to adapt to the scenario of analyzing web pages. Consistent with Claude Code and Cline's approach, AIPex does not embed and store your web pages, but uses optimized search to let the large model judge which elements are needed. It's neither passing all page content to the large model, nor using embedding-based RAG technology.

These technical innovations enable AIPex to significantly reduce computational costs and response time while maintaining high accuracy.

4. Skills Support

AIPex seamlessly integrates with Claude Agent Skills, opening unlimited possibilities for browser automation:

• Import Skills: Access thousands of pre-built skills created by the community, expanding automation capabilities
• Export Skills: Export successful AIPex workflows as reusable skills
• Skill Combination: Mix and match multiple skills to create complex automation workflows
• Ecosystem Collaboration: Benefit from the collective knowledge of the Claude ecosystem

This means you can not only use AIPex, but also leverage the entire Claude Agent Skills ecosystem, making your tasks reusable, shareable, and more efficient.

5. Intelligent Intervention

AIPex intelligently prompts users for confirmation when tasks require it, ensuring the security of critical and sensitive operations such as payments and confirmations.

6. Targeted User Scenarios

AIPex can understand web pages and user actions, so it has made targeted optimizations for specific scenarios, such as writing user guide documents ("How to create a domain on Vercel?").

Previously, if you wanted to write user documentation for your system, you needed to:

• Return to the user perspective, ensuring the documentation doesn't include technical terms
• Manually record each step and write descriptive documentation for each step
• Manually screenshot each step and add key annotations
• Organize and format the documentation for each step, and finally form the document

But now, you just need to open AIPex's user guide function, record your operations, and AIPex will automatically generate user guide documentation for you.

This efficiency improvement is revolutionary. As a human, you no longer need to focus on formatting, user perspective, or technical terms—AIPex handles all of that for you. You only need to care about the final product and can update it at any time. There are many similar niche scenarios like "writing user guides," such as end-to-end testing and recording product demos. AIPex can provide better solutions for these niche scenarios—stay tuned.

How AIPex Was Born

Initially, I just wanted to build a raycast-like tool within the browser that could be summoned from anywhere, helping me switch tabs (similar to Arc browser's Command + T shortcut, selecting tabs to switch), organize tab pages (I often need to handle 40+ tabs, manual organization is very troublesome), and summon AI assistant from anywhere (whether sending emails, tweeting, or asking questions). So I developed the first version of AIPex. This version could optimize the multi-tab problems I encountered and could ask AI questions on some pages, but I felt it wasn't cool enough.

At this time last year, Anthropic proposed the Computer Use Operator concept, followed by Browser Use proposing the AI browser automation concept. With technological development, mainly the development of tool use and MCP, some Chrome MCPs appeared, such as mcp-chrome, playwright-mcp, browserMCP, and devtools-mcp projects. I tried them in Cursor, and the biggest problem was that they all used headless browsers, which couldn't reuse user login states, and couldn't even help me post on Xiaohongshu without intervention. Actually, this separation of MCP client and servers also has context waste problems—Cursor couldn't perform targeted context optimization.

So I wanted to build a Chrome extension that could be used directly in the browser, reuse user login states, control browser behavior with natural language, and perform targeted context optimization for browsers. Before this, I actually didn't understand what MCP was, what tool use was, or what Agent Loop was. After wrestling with Cursor for a week, I had the first version of AIPex, covering 80+ browser tools. At that time, I open-sourced the AIPex code and recorded the first demo video "Help me use Google to research MCP." AIPex would open Google, enter MCP, click search, further click into sub-links for research, and finally generate a report about MCP.

I shared this demo with my leaders, colleagues, and friends, and they were all very interested, wanting to understand what was done here. Initially, I treated AIPex as a toy made in my spare time. Glace was the first friend to contribute. He has unlimited enthusiasm and ideas for AIPex, and he hopes to use AIPex's capabilities to solve actual problems encountered at work, such as writing user documentation and interface end-to-end testing. I would communicate product forms and requirements with him. What we had in common was that although we were both TypeScript beginners, we both had absolute faith in Cursor's code quality. Glace's high energy further influenced me, making me realize that AIPex is interesting and valuable, which led me to further optimize AIPex. Without this person, AIPex wouldn't be where it is today.

As more colleagues and friends learned about AIPex, we received more requirements and modification suggestions. The more serious issue was that the first version's UI was quite ugly, almost mixing native components with third-party components, resulting in an unattractive, inconsistent UI style, and some bugs. Ken completely rebuilt AIPex's UI using AI elements during National Day, and this code change accounted for 1/2 of the codebase at that time. The final result was stunning, giving us the current AIPex UI. Later, when Claude Agent Skills appeared, Ken also replicated the Skill 1:1 and integrated it into AIPex within a week. Skills can use prompt + scripts to record the successful execution process, and the next time you use it, AIPex can perform more consistently and quickly based on the Skill.

Currently, k8s/golang/supabase/tidb contributor 卡神 has also joined us. After studying mainstream AI Agent implementations in the market such as gemini-cli, openai-agents-sdk, and spring-ai, he refactored our open source repository. AIPex's open source code has become more understandable, more standardized, and easier to maintain. 卡神 will continue to help us with open source community operations.

We are still a small team of 4 people. Glace and I are primarily responsible for the product, while Ken is a senior frontend developer, which is crucial for AIPex as a rich frontend model Agent. 卡神 is a senior open source contributor, which is very important for our project's open source roadmap and community operations. Currently, we are all working on AIPex in our spare time. On the product side, our goal is to continue developing more suitable vertical scenarios, such as recording product demos. On the marketing side, I'm trying to optimize SEO. I've actually been researching this for a year and have achieved some results, but haven't seen great effects for AIPex yet. Our current goal is to achieve stable and substantial MRR. If you have any collaboration intentions, you can also contact us at https://www.claudechrome.com/contact.