> ## Documentation Index
> Fetch the complete documentation index at: https://docs.perceo.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Overview

> Fleet manager for computer-use agents running on isolated VMs.

Archfleet is a fleet manager for computer-use agents. Draw a workflow, run it on
an isolated libvirt VM driven by [Agent S](https://github.com/simular-ai/Agent-S),
and take over via XRDP when the agent gets stuck.

<Info>
  Archfleet is the newest product in the suite and is in active development. This
  page describes the shape of the tool; install and configuration docs land with
  the first tagged release. Follow
  [perceo-ai/archfleet](https://github.com/perceo-ai/archfleet) for progress.
</Info>

## The problem it solves

Computer-use agents need a real desktop — a window manager, a browser with a
profile, a file picker, a login session. Giving one your actual machine is not
something you want to do twice.

Archfleet gives each run its own VM instead. The agent gets a full desktop; you
keep yours.

## How it works

<Steps>
  <Step title="Draw the workflow">
    Define what the agent should accomplish as a workflow rather than a single
    prompt, so a long task has structure the runner can follow and you can inspect.
  </Step>

  <Step title="Run it on an isolated VM">
    Each run gets its own libvirt guest. The agent has a real desktop with real
    applications, and nothing it does touches the host.
  </Step>

  <Step title="Driven by Agent S">
    Agent S is the computer-use harness — it does the perceiving and acting inside
    the guest. Archfleet owns everything around it: VM lifecycle, workflow
    definitions, scheduling across the fleet, and run history.
  </Step>

  <Step title="Take over when it stalls">
    Connect to the guest over XRDP and drive it yourself. Computer-use agents get
    stuck, and the useful answer is a working remote session, not a failed run.
  </Step>
</Steps>

## Where it sits in the suite

Archductor produces the change. Archfleet is where you find out whether the change
actually works when a human-shaped user drives the product.

The same isolation principle applies in both: one task, one environment, one
reviewable result. In Archductor that unit is a workspace; in Archfleet it is a run.

## Requirements

Archfleet runs on your own hypervisor host. You will need libvirt/KVM with enough
capacity for the guests you intend to run in parallel, guest images with a desktop
environment and XRDP, and model access for Agent S.

<Note>
  Exact version requirements, the install path, and the workflow schema are
  pinned with the first release. Nothing here is a supported API yet.
</Note>