# xlang-ai/OSWorld

[NeurIPS 2024] OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments

Repository: https://github.com/xlang-ai/OSWorld
Canonical: https://ross.abutalabs.com/products/osworld
Homepage: https://os-world.github.io
Language: Python
License: Apache-2.0
License Family: permissive
Topics: agent, artificial-intelligence, benchmark, multimodal, reinforcement-learning, rpa, code-generation, language-model, cli, gui, natural-language-processing, large-action-model, llm, vlm
Last push: 2026-08-21T09:11:07+00:00

## Health v2 (maintenance only)
Score: 62/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 98, release rhythm 8, longevity 75
- inputs: {"age_days": 1053, "days_push": 12, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 3108, forks 526 (observed 2026-08-28T04:07:43.806157+00:00)

## What it is
OSWorld is a benchmark and scalable real computer environment for evaluating multimodal agents on open-ended computer tasks across operating systems, with 369 real-world tasks and reproducible setup/evaluation scripts. It supports task setup, execution-based evaluation, and interactive learning in VMs hosted via VMware, VirtualBox, Docker, or cloud providers.

## Use cases
- benchmark multimodal llm agents on real computer tasks
- evaluate gui agents in a real desktop environment
- compare computer-use agent performance across models
- set up an interactive environment for agent reinforcement learning
- test agents on tasks involving arbitrary desktop applications
- run reproducible evaluations of os automation agents

## When to choose
- you need a standardized, reproducible benchmark for computer-use or GUI agents
- you want execution-based evaluation rather than static datasets
- you need a real OS environment spanning multiple applications for agent testing

## When to avoid
- you only need a lightweight text-only agent benchmark without OS interaction
- you cannot provision virtual machines or cloud instances for evaluation
- you need a production agent framework rather than an evaluation benchmark

## Facets
- artifact type: dataset
- maturity: active
- function: benchmarking, agent-framework, testing
- domain: artificial-intelligence, large-language-models, developer-tools
- platform: python, cloud
- tags: multimodal-agents, computer-use, gui-agents, evaluation-benchmark, virtual-machines, reinforcement-learning-environment, ai-agents, linux, docker

## Member repositories
- xlang-ai/OSWorld (main) score 62

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:07:43.806157+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T07:26:52.262449+00:00, confidence not recorded.
  - readme: https://github.com/xlang-ai/OSWorld (fetched 2026-08-28T04:07:43.806157+00:00, sha c58e6753c168)
  - homepage: https://os-world.github.io (fetched 2026-08-29T09:42:00.520382+00:00, sha 2137f76516a6)
- Data as of 2026-08-30T08:39:29.467469+00:00.