AI NewsDev toolsAnnouncement

AI Homepage writes a new tab from your browsing history, 316 stars

AI Homepage is an MIT-licensed Chrome extension, 316 stars on day one, that replaces the new tab with a page a Claude agent writes from your own browsing history and the pages it opens signed in as you.

AI News

Editorial3 min read

LinkedInX
ThariqS/ai-newtab GitHub repository card

Image: GitHub

Why it mattersA reader who installs it gets a personalised front page without a server in between, pays Anthropic for the model call themselves, and can watch exactly which pages the agent read and copied from.

A Chrome new tab normally shows a search box and the sites you visit most. A new open-source extension replaces it with a page a Claude agent writes for you each morning, after reading your browser history and opening the pages it ranks highest while signed in as you.

AI Homepage was published by a developer who writes under the GitHub name ThariqS. It is MIT-licensed and crossed 316 stars inside twelve hours of the first commit.

How a build runs

The extension drives a Claude Managed Agents session from a background worker, with no server in between. The session has two custom tools provided by the extension: getHistory returns a ranked summary of the user's most-visited domains, and getPageHtml opens up to eight of those pages in a background window and mounts the cleaned-up HTML as files in the agent's sandbox. Inside the sandbox the agent uses grep, read and write to pull headlines, posts and links out of those files, then writes one React component the new tab renders.

Search engines, webmail, banking and logged-out marketing pages are skipped. Anything the agent cannot open is reported in the progress view, not hidden. The README states that every link on the finished page has to come from a page the agent actually read, and that it is told never to invent a URL.

What the user can change

A box in Settings holds free-form instructions that go into every build and override the agent's defaults. The README's own table lists five looks a reader can type in: a serif newspaper with a lead and a column of briefs, a quiet one-column reading list with one line per link, a dense dashboard with GitHub and Hacker News first, a green-on-black BBS terminal, and a rules box that skips sports and politics and always shows weather and the next three calendar events. The model picker offers Claude Sonnet 5.5 as default, Claude Haiku 5.5 as the fastest and cheapest, and Claude Opus 5.5 as the most careful editor. A rebuild runs on a 6, 12, 24 or 48 hour interval, and only starts after the first manual build.

Privacy and cost

Pages are loaded in the user's own signed-in browser, which the README notes can include private content. Each build sends a browsing history summary and the HTML of the eight pages to the Anthropic API with the user's own API key. The agent session and every uploaded page are deleted when a build ends. The API key stays in the extension's local storage, which the project's status note calls out: anyone with access to the browser profile can read it, so the extension is a self-run tool rather than one to publish to the Chrome Web Store.

Each build is a single Managed Agents session billed to the user's Anthropic key, so the cost comes from the model chosen and the length of what it reads. The repository is split across three packages: the host-agnostic agent loop in packages/agent-core, the MV3 extension in apps/extension built with WXT and React, and a Bun harness that runs the agent outside Chrome and loads the built extension into headless Chrome without any real API calls.

A product manager, writer or founder who already pays Anthropic for a chat subscription can run this against their own key for the price of a model call per build, and watch, in the sidebar, exactly which pages the agent chose to read before writing their homepage.

Source

Primary source: ThariqS/ai-newtab on GitHub, MIT.

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project