OpenViSTAI · Agents · Blockchain · Crypto
Menu
Agent Evaluation & Observability

Osworld

Osworld is an open-source Python research benchmark and virtual computer environment for evaluating multimodal agents on open-ended desktop tasks.

Independent OpenViST project profile, based on linked sources.

Osworld — OpenViST project profile in Agent Evaluation & Observability.
OpenViST project illustration. See the linked sources for product details.
Understand the project

Practical overview

When it fits

Use Osworld when an agent must be evaluated through tasks performed in a real computer environment, rather than through a text-only benchmark.

Who it is for

  • Researchers benchmarking multimodal computer-use agents
  • Developers comparing models or testing changes to GUI-operating agents
  • Teams reviewing agent task completion across computer workflows

Documented uses

  • Run agents against standardized GUI benchmark tasks.
  • Compare task success across models and domains.
  • Review a specific task manually and inspect its execution evidence.
  • Track regression in agents that operate computer environments.
Before you choose

Requirements & limits

The public GitHub repository is licensed under Apache-2.0.

  • Choose VMware or VirtualBox for a non-virtualized desktop or bare-metal host; use the documented Docker route for virtualized servers.
  • Docker evaluation is intended for hosts with KVM support; the project notes that macOS generally lacks KVM support.
  • Baseline runs require an API key for the model provider. Some benchmark tasks additionally require Google OAuth configuration or proxy settings.
  • The repository documents Python 3.10 or later and installation of its requirements.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Set up OSWorld on a desktop or bare-metal host

Install the repository dependencies and configure a VMware-based environment.

Before you start

  • Python 3.10 or later
  • Git and pip
  • VMware Workstation Pro, or VMware Fusion on systems with Apple chips
  • The vmrun command configured

Steps

  1. Clone the OSWorld repository and change into its directory.

  2. Install the dependencies with pip install -r requirements.txt.

  3. Install VMware and configure vmrun.

  4. Run vmrun -T ws list to check that VMware is configured.

Expected outcome

The documented setup is ready for OSWorld to download the required virtual machines and configure the environment.

Guide limitations

  • The guide distinguishes non-virtualized hosts from virtualized platforms; consult the Docker instructions for the latter.
  • VirtualBox is also supported, but parallelism and macOS on Apple chips may not be well-supported.
Source-based practical guide

Configure Docker as the OSWorld environment provider

Prepare the Docker provider for experiments on a server with KVM support.

Before you start

  • A Linux host with Docker installed
  • KVM capability is recommended

Steps

  1. On Linux, check KVM capability with egrep -c '(vmx|svm)' /proc/cpuinfo.

  2. Install Docker Desktop or Docker Engine as appropriate for the host.

  3. When initializing DesktopEnv, set provider_name to docker and os_type to Ubuntu or Windows according to the virtual machine operating system.

Expected outcome

The documented configuration selects Docker to host the experiment's virtual machine.

Guide limitations

  • macOS hosts generally do not support KVM; the project advises using VMware on macOS.
  • Abnormally interrupted experiments may leave Docker containers that can affect system performance.
Evidence and transparency

Sources & review

Source record details
Source availability
Active
Sources checked
30 September 2026
Original website title
xlang-ai/OSWorld · Official homepage title · retained in its source language