Software Development & Delivery · Public

OSWorld

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

Best fitUsers who want source-backed project understanding before installing it.

Check whether this project matches your task before installing it.

What it can doskill, recipe, host_instruction, eval, preflight

Review the portable capability path.

Before continuingVerify in a sandbox

Do not treat a preview pack as a proven local install.

GitHub snapshot2.9k stars

478 forks · 94 contributors

Doramagic.ai Last verification date: 2026-06-14 Verification method: source evidence, semantic profile, public page gate, and static build acceptance.

Official first step Read manual preview Source repository

Publication status · 2026-06-14

What is OSWorld?

[NeurIPS 2024] OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments
Best fit: Users who want source-backed project understanding before installing it.
Not for: Not for users who want to skip sandbox verification or cannot accept configuration, permission, or maintenance overhead.
Capability added to an AI workflow: skill, recipe, host_instruction, eval, preflight
First safe verification step: Verify the smallest path in an isolated environment and keep a rollback path.
Verification state: source, Quick Start, and sandbox install checks are recorded as passed.
Top risk: May increase setup, validation, or first-run risk for the user.
Evidence base: https://github.com/xlang-ai/OSWorld, https://github.com/xlang-ai/OSWorld#readme, Human Manual, Pitfall Log

Quick decision

Use this section to decide whether the project is worth a deeper read.

Best forUsers who want source-backed project understanding before installing it.

Match the project to your task before installing it.

Capabilityskill, recipe, host_instruction, eval, preflight

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

Repositoryxlang-ai/OSWorld

2.9k stars · 478 forks

What it can do

Translate the upstream project into concrete capabilities the user can judge before installing.

OSWorld Overview & System Architecture

Related topics: VM Providers, Desktop Environment & Server, Agent Implementations, Evaluators & Benchmark Tasks, Deployment, Workflows & Common Failure Modes

Source: https://github.com/xlang-ai/OSWorld / Human Manual

VM Providers, Desktop Environment & Server

Related topics: OSWorld Overview & System Architecture, Agent Implementations, Evaluators & Benchmark Tasks, Deployment, Workflows & Common Failure Modes

Source: https://github.com/xlang-ai/OSWorld / Human Manual

Agent Implementations, Evaluators & Benchmark Tasks

Related topics: OSWorld Overview & System Architecture, VM Providers, Desktop Environment & Server, Deployment, Workflows & Common Failure Modes

Source: https://github.com/xlang-ai/OSWorld / Human Manual

Deployment, Workflows & Common Failure Modes

Related topics: OSWorld Overview & System Architecture, VM Providers, Desktop Environment & Server, Agent Implementations, Evaluators & Benchmark Tasks

Source: https://github.com/xlang-ai/OSWorld / Human Manual

Doramagic Pitfall Log

Source-linked risks stay visible on the manual page so the preview does not read like a recommendation.

Source: Doramagic discovery, validation, and Project Pack records

Sources: https://github.com/xlang-ai/OSWorld, Human Manual, Project Pack evidence, and downstream validation signals.

Community Discussion Evidence

Project-level external discussion stays visible on the detail page, not only inside the manual.

Stars2.9k stars

Forks478 forks

Contributors94 contributors

Licenseunknown

Community Discussion Evidence

8 source-linked items

Review these external discussions before using OSWorld with real data or production workflows. They are review inputs, not standalone proof that the project is production-ready.

01
Pixel-blind CLI agent scores 77.9% on OSWorld test_all (vs 64.3% vision)
github / github_issue
02
Container starts but Chrome DevTools port returns 400, even with clean h
github / github_issue
03
Proposal: trace diagnostics for computer-use agent failures
github / github_issue
04
Guest VM shows a Snap Store "software updates available" popup on reset,
github / github_issue
05
Feasible-task evaluators return reward=1 without verifying the task was
github / github_issue
06
v0.1.16
github / github_release
07
v0.1.0
github / github_release
08
Capability evidence risk requires verification
GitHub / issue

How to start

Only source-backed commands are shown here. Verify them in an isolated environment first.

Try the prompt first

Test the workflow without installing the upstream project.

preview

Read the Human Manual

Understand inputs, outputs, limits, and failure modes.

manual

Take context to your AI host

Use the compiled assets in your preferred AI environment.

context

Run sandbox verification

Confirm install commands and rollback before using a primary environment.

verify

pip install osworld

Official start command · https://github.com/xlang-ai/OSWorld#readme · verified: yes

Human Manual

The English page must expose the real manual, not a short placeholder.

8+ sections · Human Manual

OSWorld Manual

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

Open the full manual

https://github.com/xlang-ai/OSWorld Project Manual
Table of Contents
OSWorld Overview & System Architecture
Related Pages
Purpose and Scope
High-Level Architecture
Core Components
Environment Providers and the Desktop Server

OSWorld Overview & System Architecture

Related topics: VM Providers, Desktop Environment & Server, Agent Implementations, Evaluators & Benchmark Tasks, Deployment, Workflows & Common Failure Modes

Source: https://github.com/xlang-ai/OSWorld / Human Manual

VM Providers, Desktop Environment & Server

Related topics: OSWorld Overview & System Architecture, Agent Implementations, Evaluators & Benchmark Tasks, Deployment, Workflows & Common Failure Modes

Source: https://github.com/xlang-ai/OSWorld / Human Manual

Agent Implementations, Evaluators & Benchmark Tasks

Related topics: OSWorld Overview & System Architecture, VM Providers, Desktop Environment & Server, Deployment, Workflows & Common Failure Modes

Source: https://github.com/xlang-ai/OSWorld / Human Manual

Deployment, Workflows & Common Failure Modes

Related topics: OSWorld Overview & System Architecture, VM Providers, Desktop Environment & Server, Agent Implementations, Evaluators & Benchmark Tasks

Source: https://github.com/xlang-ai/OSWorld / Human Manual

Doramagic Pitfall Log

Source-linked risks stay visible on the manual page so the preview does not read like a recommendation.

Source: Doramagic discovery, validation, and Project Pack records

AI Context Pack and portable assets

After deciding to continue, take the project context into your own AI host.

Complete pack plus user-owned assets

These files are planning and verification assets for Claude Code, Codex, Gemini, Cursor, ChatGPT, and other AI hosts.

Download complete pack Read Human Manual

BundleComplete Project Pack AssetAI Context Pack AssetBoundary & Risk Card AssetHuman Manual AssetPitfall Log AssetPrompt Preview AssetQuick Start EvidenceREPO_INSPECTION.json

Preflight checks

Treat this page as a planning asset, not proof that your local environment is ready.

The manual is generated from source-linked project files and Doramagic validation signals.
Community evidence warnings stay visible instead of being converted into marketing claims.
This English page is indexable because the locale quality gate passed and explicit English index approval is enabled.
Use the upstream repository as the final authority for installation commands, license, and version-specific behavior.

Pitfall Log and verification risks

Doramagic surfaces high-risk items before users treat a candidate capability as verified.

high

Installation risk requires verification

May increase setup, validation, or first-run risk for the user.

high

Maintenance risk requires verification

May increase setup, validation, or first-run risk for the user.

high

Security or permission risk requires verification

May increase setup, validation, or first-run risk for the user.

high

Security or permission risk requires verification

May increase setup, validation, or first-run risk for the user.

high

Security or permission risk requires verification

May increase setup, validation, or first-run risk for the user.

medium

Capability evidence risk requires verification

May increase setup, validation, or first-run risk for the user.

medium

Maintenance risk requires verification

May increase setup, validation, or first-run risk for the user.

medium

Security or permission risk requires verification

May increase setup, validation, or first-run risk for the user.