Skip to content

wrapvm ships a desktop VM for agents as one Containerfile

A ten-day-old repository packages XFCE, a shared Chrome and browser and desktop skills for agents into an image that CI publishes and a Smolfile boots.

Draftship Editorial3 min read

On 2026-11-14, the wrapvm desktop VM for agents appeared among GitHub's trending AI repositories. The repository was created on 2026-11-04 and had gathered 381 stars and 18 forks in the ten days since, according to the listing. Its main language is Dockerfile.

The project's README describes it as "wrap's desktop image, as one Containerfile" that CI publishes as ghcr.io/tobi/wrapvm:desktop, plus a Smolfile that boots it. In plain terms, it is a preassembled machine image that gives an agent a graphical desktop and a real browser, rather than asking each team to wire those pieces together themselves.

What is inside the image

The README lists the components concretely:

  • Xvfb on display :1 running XFCE, viewed in a browser through voxl on port 6080, protected by a password, with wrap-desktop url printing the sign-in link.
  • One persistent Google Chrome with the Chrome DevTools Protocol on port 9222, and on 9223 across guest interfaces.
  • agent-browser for browser use, attached to that same Chrome.
  • agent-desktop for computer use, connected through AT-SPI.
  • desktop and browser skills installed in ~/.agents/skills, which point agents at those two command-line tools' own guides.
  • PulseAudio with a null sink for sound, streamed by the viewer and recordable with ffmpeg. ffmpeg is installed.
  • The shared Chrome set as the default browser via xdg-open, XFCE and $BROWSER, with its sandbox on.
  • A shell prompt that reads [vm: ] under smolvm and [container: ] inside a container.

How it is built and run

The Containerfile is derived from tobi/wrap's Containerfile.desktop at commit c562a7d, plus wrapvm's own additions, each marked wrapvm:. The base image is ghcr.io/tobi/wrap:latest, and wrap's CLI is kept. A Dockerfile symlink exists so docker build . works.

There are two ways to run it. ./wrapvm uses smolvm to create one machine per directory, start it and drop into a zsh shell, printing the viewer's sign-in link; wrapvm -- CMD runs a command instead. The container route is docker run -d -p 127.0.0.1:6080:6080 ghcr.io/tobi/wrapvm:desktop, followed by docker exec wrap-desktop url to get the link. WRAPVM_DESKTOP_PORT selects a different host port when the machine is created.

Browser configuration is handled through AGENT_BROWSER_CONFIG, set to /home/user/workspace/agent-browser.json everywhere. If that file is absent, wrap-desktop start and each login shell write {"headed": true, "autoConnect": true}. An existing file is never overwritten, and this path replaces ~/.agent-browser/config.json. For development, make test runs a podman build and the test suite, with CONTAINER_ENGINE=docker as an alternative, and make image.tar plus WRAPVM_IMAGE boots a local build under smolvm.

What it changes for people building agents

The source confirms that an agent working inside this image has two named interfaces: a browser it can drive over CDP and a desktop it can drive over AT-SPI. What it suggests, though the README does not state it directly, is that teams testing computer-use agents get a repeatable environment instead of configuring a display server, Chrome flags and accessibility on their own laptops. The [vm: ] and [container: ] prompt prefixes point in the same direction: making it obvious to a human, and possibly to an agent reading terminal output, which context it is in.

Limits and open questions

The README does not name a licence, describe network policy contents beyond mentioning that the Smolfile has one, or say how credentials are handled. It does not list supported agent frameworks or give performance figures. The star and fork counts are a snapshot from 2026-11-14 and will move. Chrome's debugging port on 9222 is documented as reachable, but the source does not say whether it is restricted to the host, which matters if the machine is ever exposed beyond a local port mapping.

Building apps with AI

Most people building with AI are shipping web apps, not provisioning desktops for agents. If your work does involve putting a model inside a product, Draftship's notes on adding AI to your app cover the app-side concerns. Draftship is not affiliated with wrapvm, and nothing in the announcement suggests the two are connected.

Primary sources

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.