Agweb: A WebKit Browser AI Agents Control from the Command Line

Agweb is a WebKit browser for AI agents controlled from the command line. It shows multiple agents’ tabs in one window and requires macOS 26 or later.

· 5 min read · 4 comments

On November 14, 2026, Agweb was introduced on Hacker News. It’s a browser for coding AI agents that requires macOS 26 or later. Its creator offers an alternative to Playwright Chrome windows that steal focus and headless browsers that make failures harder to diagnose. Agweb shows multiple agents at work in one window without bringing it to the foreground.

How agents use the browser

Agweb is built on WebKit, the engine behind Safari. An agent controls the browser from the command line: it opens a page, gets a snapshot of its structure, finds elements, and takes action. Where possible, the commands resemble playwright-cli, so an agent familiar with it can use the same approach.

Each agent gets a workspace with its own tabs. An overview panel shows every agent’s tabs and most recent command. This makes it possible to monitor parallel tasks without switching between windows. Dialogs, file selection, and downloads are handled through commands, without bringing the window to the foreground.

In a page snapshot, elements get IDs that remain valid as long as the elements stay on the page. You can request only changes to the snapshot or search for a specific element. After a command, Agweb reports what happened: for example, a menu opened, data started loading, or a console error appeared. The idea is that the agent needs a fresh snapshot less often and therefore has less text to process.

Clicks and key presses use macOS events; the browser also supports hovering and dragging. If a person needs to step in—for a login, a multi-factor authentication code, or a CAPTCHA—the ask command hands the tab over to them. A named profile preserves authentication for multiple agents. A separate command can enter a secret into a form without exposing it in the response to the agent.

Installation, resources, and limitations

The app is distributed as a DMG. Its menu lets you install the command-line tool and an Agent Skill for Claude Code and Codex. Building from source requires Xcode 26 or later. The creator has not announced versions for other operating systems.

The creator compared memory use with five inactive pages: 590 MB for Agweb, 730 MB for Headless Chrome, and 1.6 GB for Playwright Chrome. In his testing, five open pages in Agweb used 0% CPU on his Mac: inactive tabs are first throttled, then suspended. These are results from his device, not a promise of the same resource use on every site.

The ad and tracker blocker uses uBlock Origin lists, which update every four days. It can be turned off for an individual workspace. The browser can also record a tab as an MP4, including pointer movement and clicks—a useful way to show a sequence of actions.

Agweb runs only on WebKit: you can’t connect to it through the Chrome DevTools Protocol or use Chrome extensions. Sites built only for Chrome may behave differently. Response mocking works for fetch and XHR; information about image, script, and stylesheet requests is limited to their URLs. The blocker does not run uBlock Origin scripts, so some ads and requests to disable blocking remain. Where Agweb falls short, the creator suggests using Playwright.

What this means for web service teams

When a developer asks an agent to perform actions on a site, it matters to check not just the click but its result. Agweb shows tabs as the agent works and reports page changes, requests, and errors after each command. When an agent goes through a form or a sequence of screens, this makes it easier to spot the step where behavior diverged from what was expected.

Still, testing in WebKit does not replace testing the same flow in Chrome. For tasks involving advertising or analytics, the blocker can be disabled in the relevant workspace; otherwise, it will affect the page behavior you observe. A shared profile makes parallel work easier, but the team should decide which agents need access to the same authenticated session.

A recording can help a team discuss a problem without trying to reconstruct every step from memory. But it shows the agent going through the flow. Whether that flow makes sense to a person needs separate testing.

My view: watching beats guessing

I like the idea of a browser that lets you see what an agent does and what follows each command. When working with a form, one successful click isn’t enough: I want to see what happens after submission, what an error looks like, and what step is available next. A recording or snapshot of changes like this makes conversations with developers more concrete.

But I wouldn’t confuse an agent’s ability to get through a flow with an assessment of its quality. An agent can test options and gather observations faster; the team still has to understand user expectations and discuss the trade-offs itself. To me, Agweb is a testing tool, not a substitute for research and design.

Sources

Where the news comes from. The text is a retelling in the author’s own words; the facts come from the source, the opinion is the author’s.

  1. Show HN: Agweb – a browser for agents that stays out of your way github.com

About the author

Tетяна Зубко

Senior UX designer

I design web services and help teams align mockups with the product’s real-world logic. I’m happy to explain how to prepare a handoff so nobody has to guess what happens when a form errors.

All posts by the author

Leave a comment

Not published. We’ll send a link to confirm your comment.

No links or ads. Your first comment is published after review. By sending a comment you agree to the privacy notice.

  1. One thing worth testing is whether WebKit’s rendering differences from Chromium matter for the services agents are validating; a visible run helps debug failures, but it doesn’t guarantee the same behavior users see in Chrome.

    1. That’s an important caveat: Agweb’s visible run can make failures easier to diagnose, but it’s still worth checking critical flows in Chromium too. Are you thinking mainly about layout differences, or behavior such as form controls and dialogs?

  2. The overview panel showing each agent’s latest command sounds useful for spotting where a task got stuck. Can it also show a short history of actions, or do you need to inspect the agent’s own logs to understand what happened?

    1. The description only mentions each agent’s most recent command, not a built-in action history, so I’d check the agent’s logs to reconstruct what happened. The overview should help you spot where to look, but it may not provide the full trail itself.