Sulat.com / Writing on AITuesday, November 17, 2026
Agent Engineering/17 November 2026

agent-browser.dev relaunches with agent-desktop alongside the browser CLI

Chris Tate announced the new agent-browser.dev on X, crediting James Clements with the design. It puts agent-desktop beside the browser CLI. The post isn't a release, and the desktop docs have been live since 30 October.

Glossy black 3D Vercel triangle between a glowing browser window and a desktop app window, each with a cursor linked to the logo.

At a glance

  • Chris Tate announced the redesigned agent-browser.dev on X on 17 November NZ time, saying it features agent-browser and agent-desktop, and credited James Clements with the design.
  • The homepage line reads "Give your agent the tools to browse, use apps, and check its own work," and lists screenshots, video at up to 60 frames a second, Core Web Vitals checks, separate sessions and mocked API responses.
  • agent-desktop reads a desktop app's accessibility tree and returns a ref for each element, according to its README, and runs on macOS 13 or later, Windows 10 1809 or later and Ubuntu 24.04 or later.
  • Per the agent-desktop session docs, several agents get their own cursors only in a session started with the multi-agent option, and click and type fail there without an agent id.
  • The post announces a site, not a release: agent-browser 0.43.0 and agent-desktop 1.3.1 came out on 12 and 13 November UTC, and the desktop docs date from agent-desktop 1.0.2 on 30 October UTC.

agent-browser.dev has a new front page. Chris Tate posted it on X late on 16 November UTC, which was 17 November in New Zealand, and James Clements did the design. The browser CLI has company now: agent-desktop, a separate Vercel Labs project that drives desktop apps instead of web pages, sits beside it.

I covered agent-browser 0.43 and agent-desktop’s macOS sandboxes on 14 November, so this post sticks to what the site changes.

What the homepage says

The page sells the pair under one line.

Give your agent the tools to browse, use apps, and check its own work.

agent-browser.dev

Six blocks follow, and on top of the clicking and typing they cover screenshots, video at up to 60 frames a second, Core Web Vitals checks, separate sessions and mocked API responses.

Browser and desktop, side by side

agent-desktop‘s README says it gives an agent “computer use without screenshots or pixel matching”. It reads the operating system’s accessibility tree, so a Finder window comes back as a list of buttons and fields with a ref each. agent-browser‘s snapshots do the same for web pages, with a ref such as @e1 against each element. agent-desktop has no model inside, the README says, so your own agent runs the loop. It runs on macOS 13 or later, Windows 10 1809 or later and Ubuntu 24.04 or later.

Side by side: agent-browser drives a browser and reads pages as compact text; agent-desktop drives native apps and reads the OS accessibility tree.
agent-browser works on web pages and agent-desktop works on native apps; both are open source.

Several agents, one desktop

The page says several agents can share one desktop. That holds only in a session started with session start --cursor --multi-agent, per the session docs, where click and type fail without an agent id. The cursors move together while the actions take turns.

Two app windows on one desktop, each with its own agent cursor; cursors move together while clicks take turns.
In multi-agent mode each agent gets its own cursor, and their clicks take turns.

On Linux Wayland the setting saves and nothing is drawn.

What’s new and what isn’t

None of this is a new release. agent-browser 0.43.0 came out on 12 November UTC and agent-desktop 1.3.1 on 13 November UTC. Even the desktop docs predate the relaunch: the 1.0.2 release on 30 October UTC listed a docs site at agent-browser.dev/desktop.

Both tools install from npm with npm install -g agent-browser and npm install -g agent-desktop.

Sources


Discover more from Sulat AI & Tech News

Subscribe to get the latest posts sent to your email.