!
This article is a repost from I started using Claude Code.
Since Jev-ultrafast and Jev's Computer Use were introduced on X, I decided to try them out too. The browser's movements are quite fast. When I actually used it, I also started thinking about where to apply that speed.
When visiting a site for the first time and not knowing where the target information is, you have to search while looking at the page. Jev's browser operation can be used for that task. During development, there are also times when I have the AI open pages whose destination I already know, again and again.
Returning to the Money Forward screens over and over
Right now, I'm building an accounting app called "Books" that integrates with Money Forward. During development, the AI frequently opens the Money Forward screens to check them. The pages it visits overlap too, such as the fixed asset ledger and the tax filing screens.
Even if Jev operates the browser quickly, if you already know the destination, opening that URL is faster. For sites you use regularly, the target page is often still in Chrome's history. So I came up with the idea of having Jev choose a URL that matches the goal from the history.
Choosing the page to start operating from, from Chrome history
What I built is Jev Bookmarks. When the AI passes a goal like "open this page," it first looks for a page saved for that development project. If it doesn't find one, it gathers candidates from Chrome's history and has Jev choose the page to start operating from. After opening the chosen URL, jev-ultrafast operates the browser.
The history is read, when needed, from an extension installed in the Chrome I normally use. It narrows down the candidates and passes the page titles and the host names and paths of the URLs to Jev. It does not send the entire Chrome history or the query strings of URLs to the judgment for choosing an entry point.
Pages on which the operation has finished are also judged by Jev. If a page is useful for the goal, its URL is saved per project. The next time a similar request comes in, it can start from the saved page.
How saved pages, Chrome history, and Jev-based exploration are used differently.
In Books' phone book, the Money Forward fixed asset ledger and tax filing pages remain. When I want to see the same screen again during development, I can start from there. I have no record of comparing the speeds in seconds. But while using it, I clearly felt that the effort of searching for pages decreased.
If there is no suitable entry point in either the history or the saved pages, Jev Bookmarks returns "no entry point." After that, the AI can just search using Jev's browser operation. Go straight to pages you know, and explore unknown pages in the browser.
Some ideas come after you try it
While using it, I also noticed that Jev can judge multiple perspectives at once. The mechanism is also described in TypeSafe's explanation. Besides deciding on a single candidate, it could also keep three and use them in the next step. Jev Bookmarks currently has a design that chooses a single entry point.
Unless I tried Jev at least a little, I couldn't see what it could do in my own work. I experienced the browser's speed, thought that history could be used for pages I open many times, and built Jev Bookmarks. The more I use it, the more I notice the next way to use it.
Top comments (0)