AI Homepage: Let Claude read the pages you frequent every day and automatically write a new tab dedicated to youAI Homepage：让 Claude 每天读一遍你常逛的网页，自动写一张专属新标签页

The pages we open every day are highly repetitive: fixed feeds, fixed backgrounds, fixed articles. But the browser will not aggregate for you, you have to go through it yourself.

RSS and aggregator readers are the answer to the previous generation, at the cost of actively subscribing, maintaining a list of feeds, and they don't read your actual browsing behavior.

The path of AI Homepage is "let the agent read it before you": it reads it for you according to your real browsing habits, and then writes it as a page for you to sweep. You don't need a feed, it will follow your habits.

For those who do cross-border e-commerce, keep an eye on industry dynamics, and chase open source projects, the marginal value of this screen is very high - equivalent to someone turning over what you should watch for you every morning.

The extension itself is a standard Chrome MV3 extension (built with WXT + React), and the core loop is in the background worker, so you can close the page after clicking "Build" on the new tab, and the build will continue in the background. Any new tab can see the progress.

It works with a Claude Managed Agents session, which hangs two custom tools for the agent: getHistory returns a sorting summary of your most visited domains, getPageHtml opens up to 8 of them in a background window, and hangs the cleaned HTML as a file in the agent's sandbox.

Inside the sandbox, the agent uses regular grep/read/write to pull titles, posts, and links from these files and write them down as a React component - this is the homepage you see.

There is a key self-restraint in the README: every link on the page must come from a page that the agent has actually read, and it is explicitly asked never to fabricate a URL.

Pages that will be skipped include search engines, web mailboxes, banks, and all marketing pages that are not logged in; pages that cannot be opened by the agent will appear in the progress view and will not be hidden silently.

There is a free text box in the settings page, and what you write goes into each build and takes precedence over the agent's default behavior. It determines both content and style.

README itself gave five examples: writing "a morning newspaper, serif font, large print headline, a headline plus a column of newsletters" to get a big newspaper front page; writing "a quiet single column, no pictures, five worthy things with a row of reasons" to get a restrained reading list.

Write "dense dashboard: GitHub PR and issue first, then Hacker News, and finally others" to get a working cockpit; write "green black terminal wind, monospace font, like BBS in the 1980s" to get exactly the same thing.

The most practical ones are actually rule-based instructions, such as “Skip sports and politics, always show the weather and my next three schedules” - such fixed demands are faithfully implemented every day.

The model can be selected: the default Claude Sonnet 5.5 (the best balance between quality and cost), Haiku 5.5 is the fastest and most cost-effective, and Opus 5.5 is the most detailed and expensive; the new model will take effect from the next build.

This extension requires you to have your own Anthropic API Key. Each build is a real model call, and the cost is determined by the model you choose and the length of the content you read.

To generate the homepage, the extension sends a summary of your browsing history (most visited domains, number of visits, few page titles) and cleaned-up HTML for up to 8 pages to the Anthropic API. These pages are loaded in a browser where you are logged in, so they may contain private content.

Official Note: At the end of each build, the agent session and all uploaded pages are deleted; the API Key is stored locally in the extension.

But the README is also straightforward: anyone with access to this browser profile can read this key. So it's positioned as a “self-use tool” rather than a product for the Chrome Store.

The blocking list is also worth using: the hostname of "agent never reads" can be filled in line by line in the settings page, and the subdomain and * .example.com are supported; the local network addresses (localhost, private IP, .local, and .internal) are always blocked, and pages redirected to the blocked host will be discarded before the content is uploaded.

Prerequisites: Chrome or other Chromium core browser, Node.js 20 +, an Anthropic API Key.

After cloning the repository, install the dependencies and build: npx pnpm @ 10.11.1 install and then npx pnpm @ 10.11.1 build (if pnpm is already installed, pnpm install & & pnpm build directly).

Open chrome://extensions, open "Developer Mode" in the upper right corner, click "Load Extracted Extensions", and select the apps/extension/.output/chrome-mv3 directory.

Select “Hold” if Chrome asks if you want to keep the modified new tab.

The first build takes a few minutes: paste the Anthropic API Key in the new tab, click "Build today's homepage", and the progress view will step through the process of preparing the agent, reading the history, opening the page in your browser, analyzing, and writing.

This site organizes resources from the GitHub open source repository ThariqS/ai-newtab (mit license), for learning exchanges only. The project will read your browsing history and call your own Anthropic API Key. Please confirm that you accept this data flow before using it. Scenarios involving personal privacy are recommended to try it in the small browser profile first. This article describes the README published by the project. The specific behaviors and commands may change with the update of the version. Please refer to the latest instructions of the repository.