On 8 October, Thariq Shihipar shared the public repository for a Chrome extension he said he uses every day. AI Homepage replaces the browser's new-tab page with a page assembled from the places its user actually reads. Browsing supplies the material; Claude supplies the selection and presentation.
According to the project's README, a build starts with a ranked summary of visited domains. The agent chooses up to eight pages, which the extension opens and cleans in the user's signed-in browser. It reads those files and writes a React component that becomes the homepage. The sequence is clear: notice the reading habits, inspect the sources, then make a page.
Connect the story before drawing the page
The interesting part lives in the editorial prompt. Related items should be brought together across sources. An announcement and a discussion about it belong in the same story. The agent is asked to choose useful items rather than mechanically take the first few entries. That gives the page a chance to explain a developing subject.
The same prompt connects presentation to evidence. Links must come from pages the agent actually read. An image should have a real association with its item. Failed page reads must remain visible failures. These are instructions to the agent, rather than proof that every build follows them. A handsome page still needs sources that a reader can inspect.
Readers can describe the kind of page they want. The README's examples range from a newspaper to a compact work dashboard. Standing instructions can change the subjects, reading order, and visual treatment. Personalisation becomes something a person can express directly, alongside what the browser history suggests.
Freshness has a pace
The extension's rebuild code requires a first manual build before automatic rebuilding starts. Its default interval is twelve hours, with six-, twelve-, twenty-four-, and forty-eight-hour choices. A failed attempt waits another interval before trying automatically again. The refresh policy gives background work a boundary.
There is also a meaningful privacy decision. The README says browsing summaries and cleaned page content go to Anthropic using the user's API key. Because pages open in a signed-in browser, they may contain private material. It says uploaded pages and the agent session are deleted when each build ends. Those are the project's documented behaviours, not an independent privacy audit.
Turning that idea into a public publication adds another editorial decision: which material can safely and usefully become a story for other readers? The direction for a public AI desk should include checked primary sources, clearly attributed reports, and permanent articles. A personal homepage's access to signed-in pages should not become a public newsroom's source policy.
For a small experiment, give a page one explicit instruction: surface material that helps answer a current project question, retain its source links, and explain why each item belongs. Then judge it by whether it makes the next reading decision easier. That is the useful promise of an editor behind the page.